GPT-6-Astra-Pro Is Rolling Out to Pro Plans: What OpenAI's Higher Tier Actually Gives You

Open​AI gave the ChatGPT Pro tier an upgrade on October 16, and what it changed is behaviour rather than a model: GPT-6-Astra-Pro in ChatGPT now answers faster, uses plugins and tools better, and — the part that is genuinely new — runs several agents at once, each working a different piece of the same task in the same conversation, with every agent’s progress visible in the thread and each finished task openable to review what that agent did. It is rolling out to eligible Pro users in the Chat tab, and availability inside managed workspaces still depends on workspace settings and rollout eligibility (vendor-stated: Open​AI’s own ChatGPT release notes for October 16 and its own post on X the same day; “faster” is the vendor’s word and no figure accompanies it). The same twenty-eight-day shipping run this piece describes below carried a second change on the same day, aimed at the developer surfaces rather than the Pro tier: Open​AI’s Codex lead said Codex and ChatGPT Work now give more effective usage by reusing context across subagents, holding that context cached longer, and routing requests so more of them land as cache hits (executive-stated; no before-and-after cache-hit rate, allowance figure or price line was published with it). That makes six moves for this page’s subject, and the first one that adds capability to the top setting rather than changing what it costs. On October 7 Open​AI began rolling the model into the Chat tab of ChatGPT — the default conversation surface, where it takes over the “Latest” slot in the model menu — carrying a new capability called Intelligent UI: answers composed from text, graphics, tappable buttons, forms and charts, or plain prose when that is the better answer. ChatGPT Plus, Pro, Business and Enterprise accounts got it first, with Free and Go dated October 8 in the same documents and Enterprise access still subject to workspace admin settings (vendor-stated, from Open​AI’s own launch post and the help-center article updated the same day). In Chat, Open​AI says the rollout is powered by GPT-6 Sol on those paid tiers and by GPT-6 Luna on Free and Go — but the help center it published hours later says Free and Go users receive GPT-5.6 Luna and lists them as “Not included” at every thinking level, which is the rollout’s first unresolved detail and the one that decides whether free users really get the newer model at all. The second unresolved detail is what the new feature does to this page’s subject. Intelligent UI runs at the Instant, Medium, High and Extra High thinking levels, all of which GPT-6 Sol powers; Open​AI’s help page states that it is “not available with Pro effort, which remains powered by GPT-6 Astra” — so the first headline capability of the line ships on the cheaper model and skips the configuration this page has spent a month trying to pin down. Before that, the story had moved twice inside one event: at DevDay 2026 on September 29 Open​AI began rolling out dots, always-on agents that each get their own cloud computer and browser, reach more than 4,000 apps through Open​AI’s plugin ecosystem, and arrive bundled with the Pro and Business Premium plans at no extra cost, and the same keynote put Ultrafast mode, the service tier Open​AI previewed on GPT-5.6 Sol in August, on GPT-6 Astra at roughly six times the standard token rate behind a new $500-a-month Pro 500 plan, cut the included usage on the $200 Pro plan, and reopened $200 sign-ups after a twenty-day pause. Six days later the subscription got faster rather than smarter: on October 5 Open​AI’s Codex lead said the default speed across GPT-6 Astra and GPT-6.1 Sol is now roughly 50% faster — about 50 tokens per second instead of 30 — in the vendor’s own products and in partner apps that sign users in with ChatGPT, and opened a 28-day run in which each day brings either one improvement or a full reset (executive-stated; no independent measurement of the figure exists). The same day brought something that is neither a model nor a plan: textGrain, an invisible text watermark that API customers worldwide can opt into, and that ChatGPT and Codex will apply by default to eligible output in the European Union over the coming weeks. The configuration this page is about, GPT-6-Astra-Pro, is still not a model you can call, and there is still no gpt-6-astra-pro identifier on the vendor’s API. What has changed is that Open​AI’s own help center now names the engine under the top thinking level — GPT-6 Astra, the same weights anyone can call — while excluding that level from the feature Open​AI spent October 7 promoting. October 16 reverses that for one thing and leaves it standing for another: the parallel agents, the better plugin and tool use and the faster responses are documented for GPT-6 Pro, the setting the help center says is powered by GPT-6 Astra, and Intelligent UI still does not reach that setting. The other boundary is that the rollout does not reach the whole product: Open​AI’s own documents say the models behind ChatGPT Work and Codex are unchanged by the October 7 release, which leaves GPT-6.1 Sol — the newer Sol the vendor shipped on September 29 — selectable in Work and Codex and nowhere in Chat. The October 16 Codex and Work change sits inside that boundary rather than crossing it: it changes how those two surfaces spend a plan’s allowance, not which model they run. What follows separates what Open​AI’s documentation confirms about each route from what is still unquantified, and it labels every figure with where it came from.

One launch, two tiers: where GPT-6-Astra-Pro sits

OpenAI's API catalogue lists exactly one GPT-6 Astra model. The identifier is gpt-6-astra: a roughly 1.05-million-token context window (922,000-token maximum input, 128,000-token maximum output), text and image input with text output, reasoning-effort levels from low through max, and a knowledge cutoff of April 30, 2026, per the vendor's developer documentation. It is the model behind the headline numbers OpenAI published at launch — 72.6% on OSWorld 2.0 against 65.7% for GPT-5.6 Sol, 74.1% on DeepSWE v1.1, a perfect 100% on ExploitBench — all vendor-reported and none independently reproduced yet.

On the ChatGPT side, the picture splits in a way the API does not reflect. OpenAI’s launch materials describe a higher-performance configuration of GPT-6 Astra aimed at the top subscription tiers. Coverage has taken to calling it “Astra Pro”; the early-access post that triggered this article writes it as GPT-6-Astra-Pro. Whichever spelling you meet, the plan facts have been consistent across OpenAI’s launch pages and the outlets that reproduced them: GPT-6-Astra-Pro is reserved for ChatGPT’s most expensive subscriptions, and it is not available to Plus. Enterprise and Business administrators have to enable Astra for their workspace. Since October 7 the picker is documented rather than reported. OpenAI’s help center now states that GPT-6 Pro is powered by GPT-6 Astra and is available in ChatGPT on the Pro $100, Pro $200, Business and Enterprise plans, and that Astra is the Pro thinking level specifically — Instant, Medium, High and Extra High run on GPT-6 Sol instead, on Plus as well as on the plans above. Plus subscribers do meet GPT-6 in Chat from October 7, which is a wider surface than they had on September 5; what they still do not get in Chat is Astra, which reaches Plus only through ChatGPT Work and Codex, on smaller allowances. OpenAI’s September 3 line that Astra would reach “all ChatGPT Plus, Pro, Business and Enterprise users” was accurate about plans, not surfaces — and October 7 closes out the other half of it: the paid tiers now meet GPT-6 in the regular chat product, but the one they meet there is GPT-6 Sol. One more line belongs in that picture now. Since October 16 the top setting is no longer only the thing OpenAI’s new features skip: the release notes file the faster responses, the improved plugin and tool use and the parallel agents under “GPT-6 Pro” — which is exactly the setting the help center describes as powered by GPT-6 Astra — so the highest configuration of the model is the first place OpenAI has added capability since launch rather than only price and allowance. No model id, spec change or figure came with it, and the same note says availability in managed workspaces depends on workspace settings and rollout eligibility.

That makes the public record lopsided — less so since October 7, but still lopsided. There is a full spec card, an API price and a rate-limit table for gpt-6-astra. For GPT-6-Astra-Pro there is still no separate model card, no separate API identifier, no context-window figure and no OpenAI-issued benchmark. What the help center added on October 7 is an attribution rather than a specification: “GPT-6 Pro, powered by GPT-6 Astra” tells a reader which weights the top thinking level runs on — the same gpt-6-astra string the API sells to anyone — and nothing about what differs operationally. The two names a reader can meet today are therefore still different in kind: one is a model you can call, and the other is an entitlement that changes which GPT-6 Astra you get in ChatGPT. Ultrafast, announced on September 29, does not close that gap either. It is a service tier on the same gpt-6-astra model string — a request flag, not a second model — and OpenAI has not described it as the Astra Pro configuration or connected the two in any published material. What DevDay does add is the first place OpenAI names the model in a product rather than in a spec sheet: the vendor’s launch page for dots states that the agents are powered by GPT-6 Astra, the same sentence appears in its release notes, and the Help Center’s own dots article states it a third time — “Powered by GPT-6 Astra, it has its own cloud computer, works across the apps you choose to connect.” That is an attribution, not a specification, and it still does not separate a Pro configuration from the standard one — but October 7 extended the pattern to the Chat product itself, where the help center attaches the same name to the Pro setting, and it is the second time the company has attached this model to something a subscriber buys rather than something a developer calls. October 16 makes it three, and the third is different in kind: the dots launch page named the model, the October 7 help-center page named the engine under a thinking level, and the October 16 note changes what that setting does — parallel agents, better tool use, faster responses — while still publishing no identifier, no limit and no number for any of it. The developer-documentation screenshot below is the September-period capture this page has carried since launch and is deliberately left as it was — it still reads “rolling out today for enterprises in our Trusted Access Program,” which is the September 3 Daybreak wave, so read it as a dated snapshot of the spec card rather than as the current rollout state. The card’s figures are the ones OpenAI still publishes for gpt-6-astra.

The widening has not been confined to OpenAI's own surfaces. On September 5, Perplexity announced that GPT-6 Astra is now available in Perplexity Computer for its Pro and Max subscribers. Perplexity Computer is the agentic workspace where a subscriber picks an orchestrator model for long research-and-writing sessions, and the model it now offers is standard GPT-6 Astra — not the higher-performance GPT-6-Astra-Pro configuration OpenAI reserves for its own top ChatGPT tiers. Perplexity framed the addition with its own WANDR evaluation, reporting GPT-6 Astra at 0.682 and $11.98 per task, which it says is the highest score of any model it has tested; those figures are Perplexity's, not independently reproduced. The result is a second seat for the standard model: a reader can run GPT-6 Astra inside an agentic product without a ChatGPT plan or an API key, but no route that bypasses OpenAI reaches the Pro configuration. The one agent product that does look like that route is OpenAI's own: dots, described below, is an always-on agent running on GPT-6 Astra and sold with the subscription rather than per token.

The rollout, plan by plan

Access was staged, and for two days the staging was the story. On day one, September 3, GPT-6 Astra went to organizations in OpenAI’s Daybreak program and a first wave of enterprise customers, with OpenAI saying ChatGPT Plus, Pro, Business and Enterprise subscribers, the OpenAI API and AWS would follow “over the coming days,” and with no timeline for free access. The first concrete date on the third-party API side landed September 4: GPT-6 Astra began appearing in third-party API catalogues, listed at the same $10-and-$50-per-million prices and the same 1.05-million-token context window OpenAI documents for gpt-6-astra. The listings were the catalogues’ own — OpenAI’s launch materials named no third-party platforms — but they were the first evidence that the promised API widening was happening, and that the standard model was reachable at OpenAI’s list price through an API key a developer may already hold, rather than only through a direct OpenAI arrangement. On September 5 OpenAI confirmed general availability: GPT-6 Astra reached ChatGPT’s Plus, Pro, Business and Enterprise plans, ChatGPT Work, Codex and the OpenAI API, and OpenAI announced another usage-quota reset the same day. The one line the September calendar left blank — free access, with no date at all — was filled in on October 7, when OpenAI put the Free and Go tiers on the rollout from October 8 (vendor-stated; which model those tiers actually receive in Chat is the one part of it OpenAI’s two pages still disagree on, as the lede sets out). That ordering tracks the model’s safety history: GPT-6 Astra is the first OpenAI model rated “Critical” under the company’s Preparedness Framework, so the capability that earned the rating was the last thing general users got, not the first — and the October 7 release keeps the pattern in a new form, since the interface added to the free product is not the model the paid tiers reserve.

Two boundaries of that release are worth stating plainly, because the “all” in “all ChatGPT users” is doing less work than it looks. OpenAI’s launch post and its October 7 release notes both say the models powering Work and Codex are not changing with this release, and the system card published alongside it is more specific: the GPT-6 Sol and GPT-6 Luna answering ChatGPT questions from October 7 are October versions that replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT, while users reaching GPT-6 Sol and GPT-6 Luna through Codex and ChatGPT Work are still on the September versions. GPT-6.1 Sol is the sharpest edge of that: the newer Sol OpenAI shipped on September 29 is a Work and Codex option, the help center lists it separately from the Chat choices, and the September 29 announcement said it was “not yet available in Chat.” The practical consequence for anyone deciding what to build on is that the two surfaces now run different versions of the same two names, and the newest Sol is the one a reader cannot select in Chat.

The first signal that the Pro configuration would move early arrived before any of that was official. On September 3, the X account @kimmonismus — which had been posting hands-on reactions to Astra, including the observation that “it really is general intelligence … simply good at everything” — added a rollout note: “Full access to all sub plans in the coming days. GPT-6-Astra-Pro is also rolling out on pro plans.” At the time it was a single unverified social post, relayed by the AI aggregator AIHOT, and OpenAI had scheduled nothing. The September 5 announcements bore the sequencing out: full access across paid plans arrived, and OpenAI’s GA-day usage caps put GPT-6 Pro on both Pro tiers. The post was directionally right even though it was never an official schedule. One more detail belongs here, because it is about the default rather than the premium: on October 5 the Codex lead put the default speed for GPT-6 Astra and GPT-6.1 Sol at roughly 50% faster — about 50 tokens per second instead of 30 — for subscription traffic in OpenAI’s own products and in partner apps using Sign in With ChatGPT, and called the vendor’s tokenizer the most optimized it has. That figure is the vendor’s own, stated by an executive on X and measured by nobody outside, and it covers the subscription surface only — the vendor’s API changelog has no entry about speed or pricing for October 5, only an unrelated HIPAA compliance flow.

The calendar OpenAI sketched at launch closed on September 5, faster than the "over the coming days" estimate OpenAI gave on day one. OpenAI said GPT-6 Astra had reached general availability — every paid ChatGPT tier, ChatGPT Work, Codex and the OpenAI API — and the rollout reached Plus and Business standard accounts as well, not just the Pro, Enterprise and Business Premium users who went first. Usage quotas were reset again the same day. That reset extended the compensation pattern OpenAI established during the phased rollout — a banked allowance for each day a paid account went without access — into a full banked reset for every Plus, Pro and Business user. OpenAI's Codex lead said it would land by end of day Pacific time on September 5 and that accounts created or upgraded before 8 pm PT that day would receive it too; those delivery details are OpenAI's own accounting, relayed by its engineers and by users, not something a third party can audit. What a developer should actually read is the meter OpenAI published alongside it. GPT-6 Pro is capped at 200 messages a week on the $200 ChatGPT Pro plan and 50 a week on the $100 plan, the latter drawn from an allowance shared with GPT-5.6 Sol Pro; Business Standard and Business Premium allowances are lower. Those are OpenAI's published figures, not independently verified — and as of September 29 the $200 figure is out of date, because OpenAI cut it in half. OpenAI's Pro tier help page now puts a term on the old cap: subscribers who held $200 access at any point from September 22 through 10 a.m. Pacific on September 29 keep the previous allowance through October 29, 2026 and then move to the lower one at an unchanged $200 (vendor-stated). The DevDay changes are below, and one of them adds a second meter to the same plans.

Community reports fill in one practical detail that matters more than the calendar. Astra usage draws against a plan’s existing allowance rather than having its own meter, and the token math makes that allowance vanish quickly at Astra’s rates: at $10 and $50 per million tokens, a single long agent run can spend a large share of a Plus-tier allowance. Reports on Chinese-language forums describe Plus users exhausting a five-hour rolling window in a couple of conversations. Those are forum reports, not measured figures, but they are consistent with the economics below, and September 10 moved them past anecdote: OpenAI’s own capacity statement is the vendor-side version of the same observation. October 7 changes the arithmetic for the chat surface without changing it for Astra, and the help center is explicit about why: Intelligent UI has no separate usage quota, so what a plan’s GPT-6 chat usage burns is the allowance it already had, and the model doing that burning on Plus is GPT-6 Sol at $2 and $10 per million tokens rather than the $10-and-$50 flagship. The Astra side is untouched — Plus still reaches it only through ChatGPT Work and Codex, on smaller allowances — which is why the “which plan” question remains a “how much Astra can you afford to run” question even after the free tier got a date, and the separate one now is what a plan has left for the chat product afterwards — and on Plus, October 7 makes those two different models at five times the price difference, which is the closest thing to a practical answer this page can offer today.

What "higher performance" means — and the gap in the record

The single biggest unknown about GPT-6-Astra-Pro has been what OpenAI means by "higher performance," because the company has not said. The phrase could describe a higher reasoning-effort ceiling, longer or more reliable agent runs, more compute per response, or simply the same model with fewer usage guardrails on the plans that pay more. OpenAI has never published those distinctions. It has issued no GPT-6-Astra-Pro benchmark of its own, no separate context or output limits, and no statement about whether the Pro configuration changes the model's safety behavior or its reasoning transparency — the latter already a live controversy for Astra, whose written reasoning is harder to audit than GPT-5.6 Sol's.

Compare that with how OpenAI has tiered before. The GPT-5.6 family shipped as distinct named models — GPT-5.6 Sol, Terra and Luna — each with its own spec card, context window and price. GPT-6-Astra-Pro is not that. It is a configuration of an already-shipped model, and until OpenAI publishes what differs, the honest summary was that “higher-performance” is a claim about the experience on expensive plans, not a spec a developer can evaluate. The DevDay announcements get closer to an answer than anything OpenAI had published before, without giving the spec, and they do it twice. When the company described what the new $500 Pro 500 plan adds over the $200 plan, the differentiator it named was speed and usage allowance — Ultrafast access and the highest allowance on offer — not a more capable model. And when it described dots, the agents it sells on the same top plans, its launch page named the engine as GPT-6 Astra — the same model string the API sells to anyone, not a separate Pro identifier (vendor-stated). The Pro tier help page adds a third statement and the most concrete of the three: it prices the premium in allowance rather than capability, stating that Fast mode consumes a plan’s included usage at 2.5 times the standard rate and Astra Ultrafast at eight times, and that a Pro 200 subscriber who keeps the older, larger allowance still “does not upgrade your plan or add Ultrafast.” October 7 added a fourth, and it is the plainest of them: the help center now says the Pro setting is “powered by GPT-6 Astra” — the name of the model in the API catalogue, not a second identifier — on the same page that withholds the new Intelligent UI from that setting. Read those together as the company’s own gloss on what “the top tier” now means: an agent that keeps working between conversations, running on weights that are already publicly priced, sold on a faster meter, and now excluded from the one capability the company spent its latest launch promoting. All four are vendor statements about plans and products, not documents about GPT-6-Astra-Pro, and they are not the same thing — but this is still the closest OpenAI has come to describing its premium in operational terms rather than leaving it to the adjective.

The first external datapoint on the configuration landed on September 7, when the independent leaderboard SimpleBench added GPT-6 Astra to its results — and the placement says something OpenAI's spec sheet will not. SimpleBench is an everyday-reasoning benchmark, built from spatial-logic, social-situation and trick questions that have repeatedly tripped up capable models; its scores are the benchmark's own, not OpenAI's. On it, GPT-6-Astra-Pro scores 86.5%, a tenth of a point behind Claude Fable 5.1's 86.6% at the top of the board and ahead of the 83.7% human baseline, while standard GPT-6 Astra sits at 83.6%, just below that baseline. Read strictly, that is the first independent sign that "higher-performance" is a measurable difference rather than a plan-gating label: the Pro configuration clears both its standard sibling and the human baseline on the same test, and it nearly ties the current frontier doing it. It is also only one benchmark, of everyday reasoning rather than the agentic and coding work Astra's launch numbers were about, and it says nothing about what OpenAI changes operationally in the Pro configuration. But the record OpenAI left blank is no longer entirely blank.

October 7 is the first evidence on that question that comes from the product rather than the plan page, and it points away from a capability difference. OpenAI trained GPT-6 to decide the shape of a response — side-by-side comparisons, interactive diagrams, tappable buttons, forms, charts, even a small tool built for the question at hand — out of a library of native, streamable components and a compiler that processes the interface as the model generates it. That is a capability the company is clearly proud of, it is the headline of the release, and it does not run at Pro effort: OpenAI’s help page says Intelligent UI is available with GPT-6 at Instant through Extra High and “not available with Pro effort, which remains powered by GPT-6 Astra.” The levels that do carry it are all powered by GPT-6 Sol — the $2-and-$10 model the API sells to anyone. Read with the SimpleBench scores above, the two facts point the same way: the Pro configuration is real, it is measurably different on at least one everyday-reasoning test, and it is not where OpenAI puts its newest capability. The other numbers in the release are the vendor’s own and none is independently reproduced: GPT-6 Instant begins answering 44% sooner on average than GPT-5.6 Instant on questions that need web search, and in an internal evaluation GPT-6 Extra High starts answering in the same time as GPT-5.6 Medium while scoring better than GPT-5.6 Extra High. The practical details are documentation rather than benchmarks, and they are the ones a reader can act on: Intelligent UI has no separate usage quota, existing model, tool and plan allowances still apply, the capability is rolling out gradually on the web and on updated apps, and the older ChatGPT desktop apps for macOS and Windows do not support it at all. Two surfaces are out of scope entirely — the help center says the capability is not available in the Work tab and not available with Voice, so a ChatGPT Work session or a spoken conversation gets none of it. The switch is also web-only and is a softening rather than an off switch: on ChatGPT on the web, Settings → Personalization → Layout and Visuals set to Simple reduces the visual and interactive elements, though OpenAI warns some may still appear and that it is not a switch back to a classic text-only mode. October 16 pushes the other way, and it is the first real counterweight on the capability side. The Pro setting now has behaviour OpenAI has not announced for any other Chat setting: several agents working one task in parallel inside a single conversation, each with reviewable output, on top of faster answers and better plugin and tool use. That is still not a specification — no OpenAI page says how many agents run at once, how they are coordinated, what they draw from the plan allowance, or how much faster “faster” is — but it is the first update in the life of this configuration that a Pro subscriber can point at as something the top setting does and the cheaper ones do not. The other half of October 16 lands on the developer surfaces and is the opposite kind of change: Codex and ChatGPT Work now reuse context across subagents, hold context in cache longer and route requests for more cache hits, which is a claim about how far an allowance stretches rather than about what the model does (executive-stated, with no figure attached).

Why the gate is really two gates

The gating of GPT-6-Astra-Pro to the top plans has two engines, and separating them explains the rollout. The first is safety. Astra's "Critical" rating put its most advanced cyber capabilities behind locks from day one: exploit-development requests are refused in the standard configuration, and the defensive use cases live in the Daybreak Blue and Red programs rather than in the general product. Gating a heavier configuration to a smaller, higher-paying, mostly organizational audience is consistent with that posture. The October 7 release gave that ladder a published rung. The system card OpenAI published with it says the company is treating the October release of GPT-6 Sol and GPT-6 Luna as High capability in both the Cybersecurity and Biological and Chemical domains under its Preparedness Framework — High, not Critical, and neither model reaching High in AI Self-Improvement — and that it applied the same safeguard set it had documented for GPT-5.6 Sol and GPT-5.6 Luna rather than the heavier treatment Astra received (all vendor-stated; the system card is OpenAI’s own assessment and no outside party has audited it). That is the shape of the release in one line: the model rated Critical is the one behind the Pro setting, the model rated High is the one that reached a billion-plus weekly users, and the newest interface is withheld from the Critical configuration. DevDay extended the same pattern rather than relaxing it: OpenAI's description of dots is largely a description of containment — each dot works on its own isolated cloud computer with your own machine and its contents kept separate unless you connect it, a separate system called auto-review checks actions that could affect your accounts or share information against your instructions, your Custom Rules and OpenAI's safety requirements before they run, background "proactive research" is limited to read-only tools that cannot send messages, change app content or control a browser, and certain sensitive tasks stay with the user by design (all vendor-stated, and none of it independently audited).

The second engine is unit economics. GPT-6 Astra lists at 2.5 times the token price of GPT-5.6 Sol — $10 against $4 per million input tokens, $50 against $20 per million output tokens, at vendor list pricing — and its pricing carries structural penalties for the long agent runs it is designed for: prompts over 272,000 input tokens are billed at double the input and cache rates and 1.5 times the output rate, fast mode costs double again, and Ultrafast, added September 29, costs six times the standard rate. A model that expensive to serve is naturally reserved, in its heaviest configuration, for the subscribers who can carry it: ChatGPT Pro at $100 and $200 a month, and Business and Enterprise seats — and, since September 29, the $500 Pro 500 plan. The plan-side mechanics, which only became readable in the week after DevDay, point the same way: OpenAI’s Pro tier help page states that Fast mode consumes a plan’s included allowance at 2.5 times the standard rate and Astra Ultrafast at eight times, that Ultrafast is available on Pro 500 alone, and that buying credits on Pro 100 or Pro 200 does not unlock it (vendor-stated, and described there as a consumption rate rather than a guarantee of faster performance — the API’s six-times token price and the plan’s eight-times allowance burn are two different meters, not the same number). DevDay’s own billing changes point the same way. OpenAI’s release notes say dots usage will not count toward eligible Pro, Business and Enterprise plan allowances “for the next month,” after which the company says it will publish usage terms for each plan; the launch page adds that conversations with a dot do not count toward ChatGPT usage limits, while tasks the dot starts in Codex or ChatGPT Work count as usual (all vendor-stated). A bundled agent whose usage is carved out of the meter for a month is a capacity decision as much as a product one, and OpenAI has published no allowance figure to size it with. Its counterargument on cost is per-task rather than per-token — it claims the best-performing Astra configuration cuts estimated DeepSWE task cost by roughly 57% against the best GPT-5.6 Sol configuration — but that is a vendor figure, not an independent one. Cached input is where the October 16 efficiency claim gets its arithmetic, and the vendor’s own Codex rate card is the only place it is priced: GPT-6 Astra bills cached input at 25 credits per million tokens against 250 for uncached input — a tenth of the input price — and GPT-6.1 Sol at 2.5 against 50, a twentieth, with no separate cache-write charge on either (vendor rate card). A change that raises how often that row is the one you are billed on is therefore worth real money to a heavy subscriber, and OpenAI still published no measurement of it, only the mechanism. One figure from October 5 needs the same label. The roughly-50% default-speed claim is a throughput statement about the subscription path — about 50 tokens per second against 30 — relayed by an OpenAI executive on X, with no outside timing and no vendor page behind it, and no change to any rate card: the API pricing table, including the Ultrafast row, carries nothing for October 5.

September 10: OpenAI stopped selling the plan that carried the top configuration

On September 9, the OpenAI product lead who runs ChatGPT and Codex, Thibault Sottiaux, warned publicly that the company might have to stop selling its most expensive plan. The next day it did. New subscriptions to the $200-a-month ChatGPT Pro plan were paused, and the reason given was the model this article is about: demand for GPT-6 Astra. "To make sure our current users have an incredible experience and continued access to Astra, we are going to pause subscriptions to our $200 Pro plan," Sottiaux wrote. "These put the most strain on our systems and we wanted to take the smallest step that allows us to continue giving the broadest access possible." He described Astra demand as "really unprecedented" and said he had not seen anything like it before. Read the sourcing carefully: those are OpenAI statements relayed by one of its own executives, not a published capacity disclosure, and OpenAI released no figures on sign-up volume, compute headroom, or how long the pause would last. It lasted twenty days.

The scope of the pause was narrow, and worth reading precisely, because it shapes what happened next. The pause covered new sign-ups to the $200 tier only. Existing Pro subscribers kept their accounts and their Astra access, and OpenAI's other plans — the $100 plan, Plus and Go — and the API remained on sale. The $200 plan was the one OpenAI singled out as putting the most strain on its systems, which is a statement about workload shape rather than about revenue: a fixed monthly price buys an open-ended amount of long-horizon agent compute, and GPT-6 Astra is built for exactly that kind of work. OpenAI said it was adding capacity and gave no date for reopening. As of September 15 it had not reopened and the company had published no threshold for doing so; the Help Center entry stated only that sign-ups and upgrades to the $200 plan were temporarily paused, and that an existing $200 subscription which lapsed could not be restarted until the pause lifted. OpenAI later softened that: from around September 23, subscribers who held $200 access on September 10, or whose access had ended in the 30 days before, qualified for a one-time return within 30 days, exercisable by signing in and purchasing through the pricing page at chatgpt.com. That detail is OpenAI's own Help Center wording, reported by third parties and not something we have audited against the page ourselves.

September 17: a third door, and it was not a plan tier

A week after OpenAI stopped selling access to its most expensive consumer plan, it announced a new way to get GPT-6 Astra — and it did not go through the plan ladder at all. Astra for Law, unveiled September 17, is not a new model. It is GPT-6 Astra behind a legal search index, with legal-analysis and legal-drafting instructions layered on top, and tools built for United States case-law research. The index spans more than 230 million URLs covering case law, statutes, regulations, court rules and administrative decisions, with sources added daily; the case-law corpus comes from the Free Law Project's CourtListener database. OpenAI describes it as complementing, not replacing, the licensed content firms already buy from providers such as Thomson Reuters.

The launch numbers are OpenAI's, measured against OpenAI's own baseline. On 200 U.S. legal research questions drawn from the private validation set of Vals AI's Legal Research Bench, at the highest reasoning effort, OpenAI reports Astra for Law passing the overall correctness check on 54.0% of questions against 38.7% for GPT-6 Astra using web search alone — a 40% relative improvement. It also reports 24% more reference cases found on case-law-focused questions, up to 54% more relevant passages retrieved from the correct opinions, and answers roughly twice as long on average. All of that is vendor-reported on a private split, and no independent lab has reproduced it. The comparable public figure points the other way: Vals AI's own public comparison lists base GPT-6 Astra at 39.42% ±3.40 on that benchmark, essentially the baseline OpenAI is measuring against. The gap is the search index and the instructions, not a better underlying model — which is the useful thing to take from the announcement, and the reason the headline number should not be read as a jump in model capability.

Access is the part that matters here, because it is a third kind of gate. Astra for Law is being offered first to selected firms through a Trusted Access program in ChatGPT and Codex, where it appears in the model picker as “GPT-6 Astra Law.” The terms are organizational rather than retail: zero data retention, ChatGPT Enterprise usage excluded from human review by default, and a roughly 30-page agreement setting out the rest. Twenty-six partner-built plugins ship alongside it — from Relativity, Clio, iManage, Intapp, DeepJudge, Harvey, Legora and Thomson Reuters, among others — plus nine community plugins carrying 47 custom skills. OpenAI says it is working with Latham & Watkins, Sullivan & Cromwell, Ropes & Gray, Cooley and Wachtell Lipton on governance questions such as information permissions and ethical walls. An API version, gpt-6-astra-law, was described as coming soon, with no date and no price published; Harvey and Legora are named as API customers building on it. As of October 8 there is still no published date or price for that identifier, and it still resolves to no model page on OpenAI’s own developer documentation.

September 29: Ultrafast arrives on Astra, and a $500 seat above Pro

Ten days after the reopening question became the piece’s open thread, OpenAI answered it — and changed the shape of the ladder at the same time. Ultrafast mode came to GPT-6 Astra. The claim circulated first as a screenshot of OpenAI’s own documentation posted to X by the account @scaling01 on September 29; it is no longer a social post, because OpenAI’s developer documentation, its pricing table and its DevDay 2026 recap all carry it. The vendor’s Ultrafast mode guide now reads: “Ultrafast mode is the fastest service tier in the OpenAI API. It is broadly available for GPT-6 Astra, with preview access for GPT-5.6 Sol.” The GPT-6 Astra model guide carries the matching sentence — “All GPT-6 Astra users also have access to Fast mode and the new Ultrafast mode for our fastest API speeds” — and the DevDay recap lists Astra Ultrafast with availability in ChatGPT, Codex and the API. That is vendor documentation, not a leak, and when it landed it was the first time the fastest Astra path could be bought at all. Six days later OpenAI moved the baseline underneath it: from October 5 the default speed for GPT-6 Astra and GPT-6.1 Sol through the subscription is roughly 50% faster, about 50 tokens per second instead of 30, which the vendor’s own Codex lead announced on X without publishing a measurement (executive-stated). Ultrafast is still the fastest tier and still the only one priced per token; what changed is how much speed a subscriber gets without paying for it.

The tier is bought per request, not per seat, and the rate card is steep. On OpenAI's published pricing table the Ultrafast row contains exactly one model, gpt-6-astra, at $60.00 per million input tokens, $6.00 cached input, $75.00 cache writes and $300.00 output for prompts at or under 272,000 input tokens — six times the standard $10/$50 — rising to $120.00, $12.00, $150.00 and $450.00 above that line. You select it with service_tier: "ultrafast" on a request to gpt-6-astra; there is no separate model string and no separate model card. The guide adds two operational details: the tier is currently open to all API users at low rate limits, and OpenAI recommends a persistent WebSocket rather than plain HTTP for agentic work, because per-request network overhead eats into the latency the tier is bought for. Rate limits are published as 500,000 tokens per minute on usage tiers 1–3, 1,000,000 on tier 4 and 5,000,000 on tier 5, and the guide states Ultrafast supports US data residency and global processing only — it does not support EU or other non-US regional endpoints. OpenAI's own speed figure for the tier is not published in the Astra guide; the vendor says only that it is the fastest tier. The number in circulation — up to 300 output tokens per second in Codex and Work with Ultrafast enabled — comes from OpenAI's DevDay description as relayed by Engadget, and it is a claim about those two surfaces rather than a measured throughput for the API. The August preview of the tier on GPT-5.6 Sol claimed up to 750 tokens per second and up to 14× standard speed; none of the Astra figures carry that comparison, and OpenAI has not published one.

The subscription change is the other half, and it is the part that genuinely inverts the piece's earlier advice. OpenAI put Ultrafast behind a new plan. Pro 500 costs $500 a month, carries what OpenAI describes as its highest usage allowance, and includes access to Ultrafast in Codex and Work; outlets covering DevDay, including Engadget and CNET, report the plan as announced at the event rather than as a leak, and OpenAI's own recap carries a Pro 500 card with Ultrafast lettering. Since then OpenAI's Pro tier help page has filled in the mechanics, and they are all restrictions. Ultrafast is the only one of the three Pro plans that includes the tier; Pro 100 and Pro 200 list it as "Not included," and buying credits on either does not unlock it. Inside Pro 500 you select it in the model picker, it draws on the plan's included usage first and on purchased credits only after that allowance is gone, and it burns that allowance at eight times the standard rate. Keeping a grandfathered Pro 200 allowance does not change any of this — OpenAI's own wording is that it "does not upgrade your plan or add Ultrafast." At the same time the $200 Pro plan got worse on paper and better on availability. Its included Codex and Work usage drops from 20× the Plus allowance to 10×, and the ChatGPT-side GPT-6 Pro message cap drops from 200 a week to 100; subscribers who held $200 access between September 22 and 10 a.m. Pacific on September 29 keep the older, larger allowance through October 29, 2026 and then move to the lower one at the same monthly price. New $200 sign-ups reopened on September 30, and Sottiaux's framing of the trade is that the subscription "will net out at half the dollar in API spend compared to the old Pro $200 plan" — his figure, published on X and reported by The Decoder, Unite.AI and others, not a number OpenAI has put on a pricing page. The five-hour usage cap is gone for good, which is the one improvement in the package.

Put those two moves together and the question this page has been circling since September 3 finally has a partial answer — which October 7 then complicated from the other side. Until now, “what does the top of the ladder actually give you” resolved to a subscription tier with no published differentiator. Since September 29, the top tier has a named one: speed, and allowance. Pro 500 is the first OpenAI consumer tier whose headline extra is latency rather than a more capable model — and that is consistent with everything else here. Ultrafast does not make GPT-6 Astra smarter; it makes the same weights answer faster for six times the token price, and inside a plan for eight times the allowance burn. If GPT-6-Astra-Pro is a capability configuration, Ultrafast is not it. If it is an access configuration, Ultrafast is the closest thing OpenAI has yet sold to a retail name for it. The October 5 default-speed change narrows that differentiator without removing it, and October 7 inverts it in one respect: the new interface goes to the models under the Pro setting rather than to the setting itself. The subscription’s standard lane now runs roughly 50% faster by OpenAI’s own account, so the paid tier is no longer the only way to buy latency on these models — it is the way to buy the most of it, at a price the vendor states and no outside test has yet measured — and, from October 7, it is the one configuration OpenAI keeps the new interface out of by design.

Dots: the first product OpenAI has shipped on GPT-6 Astra

Ultrafast was the pricing half of DevDay. The product half was dots, and it is the part that changes what a subscriber can actually get. Dots are always-on agents: each one holds a goal across sessions rather than answering a question and stopping, gets its own cloud computer and its own browser, keeps working while you are away, learns from feedback over time, and brings results back for review. OpenAI's launch page says they can reach more than 4,000 apps through the plugin ecosystem, that a dot runs inside ChatGPT as well as in Slack and Microsoft Teams with texting running as a limited beta for Pro users in the US only, and — the line that matters here — that they are "powered by GPT-6 Astra" (vendor-stated; the same weights the API sells, since no dots-specific model identifier exists).

Access and billing are the specifics worth keeping. Dots began rolling out on September 29 to Pro and Business Premium users in eligible markets, and OpenAI's launch page says the first dot is included in those plans at no extra cost; Enterprise workspaces, including Edu and Healthcare, can try the beta once an administrator enables it, and a dot has to be created on the desktop app or desktop web rather than on mobile. OpenAI's release notes add two qualifiers the launch page does not spell out: rollout is limited to users aged 18 and older in supported markets, and "Pro access excludes the European Economic Area, Switzerland, and the UK at launch." That matters more than it looks — a subscriber in the EEA or the UK can be paying for the same Pro plan OpenAI is promoting the feature with and still not get a dot, and the exclusion is stated in a release-notes line rather than on the plan comparison a buyer reads. On usage, OpenAI's own wording splits two cases: conversations with a dot do not count toward ChatGPT usage limits, while tasks the dot starts in Codex or ChatGPT Work count as usual. Those are vendor statements about a product still rolling out — the Help Center still describes access as gradual, with a wait of several days. OpenAI has published no allowance per dot, no price for a second dot and no cost per finished task, and it has published no benchmark of dots at all. Two of the guardrails it has published are more specific than the keynote, and both matter to anyone deciding whether to hand a dot real work. Sign-in is kept out of the model's context: for supported sites the model is paused while you complete a secure login form, and the saved-password flow serves credentials from a separate encrypted service, so a dot can work inside an account without the password ever reaching the conversation — with OpenAI's own caveat that a secret pasted into a message or document can still be visible. And the control that enforces action review is described as sitting outside the environment a dot can change, so a dot cannot turn off a required check; your approval can prompt a re-review but cannot override a core safety requirement (all vendor-stated, from OpenAI's dots safety pages, and none of it independently audited).

For this article's central question, dots answers it from a direction nobody had priced. GPT-6-Astra-Pro has never had a published difference from the standard model, and dots does not supply one: the agent runs on the model string anyone can call, and nothing in OpenAI's materials says a dot gets a capability configuration the API lacks. What dots does establish is that OpenAI is now willing to attach its most expensive plans' headline feature to those same weights — and to gate delivery of that feature by plan, market, age and workspace setting rather than by model. If GPT-6-Astra-Pro is an entitlement rather than a different model, dots is what the entitlement looks like in 2026: not a smarter answer but an agent with its own computer, sold with the seat. If it is a capability configuration, then OpenAI's own flagship agent product is not running it — which is its own kind of answer.

Astra Pro is still a seat, not a token — and that now cuts both ways

The consequence that matters most for builders is unchanged: GPT-6-Astra-Pro cannot be bought per token. There is still no gpt-6-astra-pro identifier in the OpenAI API, and OpenAI has given no signal that one is coming. If a team wants whatever the top Astra configuration is inside its own product, the only route remains organizational — a seat on Pro, Business or Enterprise with the model enabled by an administrator, and now the $500 Pro 500 tier above those. For API development there is still exactly one Astra model you can call, gpt-6-astra, and its September 29 addition is a flag on the request rather than a second model. The September 17 announcement adds a second identifier on the horizon, gpt-6-astra-law, still not open, not dated and not priced. Dots adds a third thing that is not an identifier at all: OpenAI describes it as a ChatGPT product, publishes no dots API and no dots model id, so the always-on agent it demonstrated at DevDay is not something a developer can call, wrap or test. The October 5 speed change does not create one either: it is a serving-layer change to traffic authenticated through a ChatGPT plan, not a model string, a tier or a price, and the vendor’s API changelog has no entry about it. The one new API surface since DevDay is textGrain, and it is a checkbox rather than a model — from October 5 API customers worldwide can opt in to watermarking for select models, off by default, as the section below sets out. October 7 adds a sixth gate without adding a seventh model string, and it is the narrowest of the lot: which ChatGPT thinking level a plan may select, which of those levels carries Intelligent UI, and whether a workspace administrator has switched GPT-6 on at all. The only trace of it in the API is a changelog line the same day recording that the chat-latest snapshot now points at the latest model available in ChatGPT for Plus, Pro, Business and Enterprise users — a pointer that moved, with no model id published and no rate-card change anywhere: the pricing table, Ultrafast row included, still reads exactly as it did on October 5. October 16 is the clearest case of the rule yet, and it runs the other way: the parallel agents, the better tool use and the faster responses exist in the Pro subscription and nowhere else — there is no API parameter, no model id and no published price for any of them, and a developer calling gpt-6-astra gets none of it. The seat-versus-token line has therefore gained its first entry on the behaviour side rather than the access side: this is the first time the plan buys something the token cannot, and OpenAI has published no conversion between them. The Codex and Work half of the same day runs the other way and is even harder for a builder to buy: cache reuse across subagents, longer cache retention and cache-hit-oriented routing are serving-and-harness changes inside OpenAI’s own clients, carrying no API parameter, no published rate change and no figure, so a developer calling the same models through the API is told nothing about whether any part of it reaches them.

What has changed is the cost of the fast path. A developer who wants Astra’s latency rather than its intelligence can now buy that by the token at $60 and $300 per million — six times the standard rate — instead of by the seat. That is a real choice where there was none before September 29, and it is the kind of decision a routing layer exists to make cheap to reverse. One API across 200+ models with provider list price passed through at 0% markup means a $10/$50 frontier model and a $4/$20 GPT-5.6 Sol sit on the same key, and automatic failover lets you send a slice of traffic at a new tier while the proven configuration carries the rest — the sane way to find out whether a 6× speed premium pays for itself on your workload before you commit a budget line to it. OrcaRouter lists gpt-6-astra at the vendor’s list price, $10 per million input tokens and $50 per million output tokens, with the same 1.05-million-token context window and the same long-context tier above 272,000 input tokens; the catalogue entry does not break out the ultrafast tier, because Ultrafast is a service-tier flag on the vendor’s own request rather than a distinct model any router can list. October 7 is worth reading the same way, because two of the three models the rollout touches are ordinary catalogue entries here: GPT-6 Sol is routed at $2 and $10 per million tokens and GPT-6 Luna at $0.10 and $0.50, both at vendor list price, and the engine under the Pro setting is the same gpt-6-astra the catalogue already carries at $10 and $50. What no router can sell is the thing the rollout actually changed — which thinking level a plan is allowed to select, and which of those levels gets the new interface. The seat-versus-token line is unchanged in the other direction too: GPT-6-Astra-Pro is the part OpenAI has not put in the API under its own name, and no routing layer can sell what the vendor does not sell per token. Dots does not change that either — there is no dots endpoint to route to, on our catalogue or anywhere else, and the honest version of the routing pitch is narrower than it first looks: you can route the model a dot runs on if you build your own loop, and you cannot route the dot. GPT-6.1 Sol is on the same list at $2 and $10 per million tokens with cached input halved to $0.10, and it is the model in this family that the October 7 release leaves out of Chat entirely — routable here, but only as the engine behind Work and Codex. October 16 is the cleanest illustration of where that line sits. The model behind the parallel-agent behaviour is, on the vendor’s own attribution, the same gpt-6-astra this catalogue carries at $10 and $50 per million tokens, and the behaviour itself is not for sale at any price on any API, so the honest split is the narrow one: one key at provider list price with automatic failover gets you the weights and lets you find out cheaply what a new serving tier or a new model is worth on your own traffic, and no routing layer can resell a harness the vendor keeps inside its own client. The cache half of October 16 is the part that does cross over, because cached input is a line item a router can pass through rather than a behaviour it has to reconstruct: on OrcaRouter cached tokens bill at the provider’s own cache rate instead of the full input price, the same 0% markup rule the rest of the card follows, and it is the row OpenAI’s change is aimed at. What no card here carries is the routing itself: whether a request lands as a cache hit depends on the prefix, the tools attached to it and the serving infrastructure in front of the model, and no vendor publishes that as a number.

What's still unverified — and what to watch

The GA settled the calendar, the September 7 SimpleBench entry gave the story its first quantitative datapoint from outside OpenAI, September 17 showed how OpenAI intends to package Astra into verticals, and September 29 gave the premium a name and gave the model a product. Four threads have landed since — a speed claim about the subscription, a watermarking switch on the API, GPT-6 itself arriving in ChatGPT’s default chat surface with Intelligent UI, and now the October 16 pair of changes — and the third one settled the most practical question of the batch, because the free tier finally has a date, which OpenAI put at October 8, and the paid tiers finally have the model named. The fourth answers two different kinds of question, and it is the only thread of the four that adds behaviour rather than availability. The October 7 thread also drew its own boundary: that release left the models behind Work and Codex alone, so GPT-6.1 Sol is still not something a Chat user can choose. Each thread also left an open question, and most of what is quantitative here is still single-source. The honest list is longer than the list of settled facts:

• Whether GPT-6-Astra-Pro is the same experience on both Pro tiers. With September 29's changes the tiers no longer share a cap: the ChatGPT-side GPT-6 Pro message limit drops to 100 a week on the $200 plan, and OpenAI describes the new $500 Pro 500 plan as carrying its highest allowance. October 16 sharpens the question instead of answering it: the release notes attach faster responses, better plugin and tool use and parallel agents to GPT-6 Pro generally, name no plan tier, publish no allowance for the agents and give no date on which the rollout finishes. "Higher-performance" is still undefined, and Business workspaces without an admin opt-in still see nothing.

• What the Ultrafast premium now costs inside a plan — and how the two meters relate. This one moved after September 29: OpenAI’s Pro tier page states that Ultrafast is exclusive to Pro 500, that it uses the plan’s included allowance first and purchased credits only after that, and that it consumes that allowance at eight times the standard rate against 2.5× for Fast mode, with the caveat that these are consumption rates rather than performance guarantees. What is still unpublished is how a plan’s eight-times allowance burn lines up with the API’s six-times token price on the same workload — two different meters, and no OpenAI page converts between them. The October 5 default-speed change does not touch either meter, because nothing about pricing or quota was announced with it. October 16 is the first time OpenAI has said what it did to make an allowance go further rather than bigger, and it came with no number attached to it: reuse context across subagents, keep it cached longer, route for more cache hits (executive-stated). Whether that stretch is measurable on a given plan is still unpublished, and it leaves the two meters above exactly as they were.

• What Ultrafast actually delivers on Astra. OpenAI’s guide calls it the fastest tier and publishes rate limits and prices but no speed multiple for this model. The 300-tokens-per-second figure in circulation is a description of Codex and Work behaviour reported from DevDay, not an API benchmark, and no independent party has measured the tier on Astra. The August GPT-5.6 Sol claims — up to 750 tokens per second, up to 14× — do not transfer. October 5 makes the comparison harder rather than easier: the vendor now says the standard lane runs at roughly 50 tokens per second, and Ultrafast’s own number for Astra is still unpublished, so the size of the premium can only be read against a baseline the same vendor states.

• What the October 5 speed claim actually measures. OpenAI’s own description is a throughput figure — roughly 50 tokens per second instead of 30, in the vendor’s products and in partner apps using Sign in With ChatGPT — and it is executive-stated, published on X with no accompanying documentation page and no outside timing. The arithmetic also does not close cleanly: 30 to 50 tokens per second is a 67% gain in throughput and a 40% cut in time per token, so “roughly 50% faster” sits between two defensible readings. Whether the API sees any of it is unstated: the vendor’s API changelog for October 5 records an unrelated HIPAA compliance flow and says nothing about serving speed.

• Whether the reopen sticks, and what it costs subscribers over time. New $200 sign-ups resumed September 30 with half the Codex and Work allowance and half the ChatGPT message cap, and the Pro tier page now dates the old allowance's end: grandfathered subscribers keep it through October 29, 2026. The "half the dollar in API spend" framing is still an executive's post rather than a published rate card, and OpenAI has not said what happens if Astra demand climbs again. Everything about this remains executive-stated rather than a published capacity disclosure.

• What “higher-performance” changes operationally — effort ceiling, run length, compute per response, and now agent parallelism. October 16 is the first partial answer, and it is qualitative: faster responses, improved plugin and tool use, several agents on one task. What is missing is every number — how much faster, how many agents, what they draw from a plan’s allowance — and any statement of whether the change is the model, the harness or the serving layer beneath both. The same day’s Codex and Work change makes that last gap concrete, because it is explicitly not the model: it is caching and request routing (executive-stated). Two changes described in the same breath as more effective usage, one on the Pro setting and one on the developer surfaces, and neither says which layer moved.

• Whether a Pro-tagged API model appears on OpenAI's own catalogue. The third-party API entries that went live on September 4 include a route literally labeled "GPT-6 Astra Pro," but the listing describes it as the same gpt-6-astra with the API's pro reasoning mode enabled at the identical price — a naming choice, not evidence that OpenAI has defined its subscription-tier Astra Pro for the API. Ultrafast, notably, is documented under its own name as a service tier rather than under a "Pro" label, which is the opposite of what a Pro-tagged API model would look like.

• How much weight the September 7 SimpleBench scores carry. GPT-6-Astra-Pro’s 86.5% and standard GPT-6 Astra’s 83.6% are the first independent numbers on any third-party leaderboard, but SimpleBench measures everyday reasoning — the agentic and coding performance behind Astra’s launch claims still has no independent check. Four newer claims are in the same position: the October 5 default-speed figure, the October 16 “faster responses” line, the October 16 cache-and-routing claim about Codex and Work, and the watermarking numbers below are all vendor-reported, and none has been reproduced outside OpenAI.

• Whether Astra for Law's 54.0% holds up. It is OpenAI's figure on a private validation split of Vals AI's Legal Research Bench, unreproduced, and measured only against OpenAI's own model with web search — no comparison against the established legal research products firms already license. The publicly visible Vals number for base GPT-6 Astra is 39.42% ±3.40, and nothing on that public board yet scores the legal configuration. When gpt-6-astra-law opens, it will need a price and a public benchmark before the vertical can be evaluated the way the rest of this article evaluates the base model.

• Whether the dots bundle survives its own first month. OpenAI's release notes carve dots usage out of eligible Pro, Business and Enterprise plan allowances "for the next month" and say usage terms per plan will be published after that; the dots documentation says only that a plan includes "an allowance for deeper work, with extended limits for the first month after launch" and that adding more dots and scaling each dot's output come later — with no allowance per dot, no second-dot price and no cost per finished task published anywhere. The EEA, Swiss and UK exclusions mean a large share of the subscribers paying for the plan cannot use the feature at all. A team weighing an always-on agent against a token budget is currently comparing a bundled plan feature to a meter, which is not yet a comparison.

• What textGrain can actually prove — and who can test it. On October 5 OpenAI opened opt-in text watermarking to API customers worldwide, said it will add an invisible watermark to eligible ChatGPT and Codex output in the European Union over the coming weeks, and began accepting applications for detector access from approved researchers and expert organizations. The performance numbers the vendor published are the only ones there are, and they are its own: at a 1% target false-positive rate its detector found watermarks in about 95% of 400-token passages against about 80% of 200-token ones for psychology-style text, with substantially lower rates where word choice is constrained, such as mathematics; replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI also says it sees no meaningful difference on its GPT-6 Astra benchmarks with watermarking on — its table lists DeepSWE v1.1 at 72.80% unwatermarked against 71.68% watermarked, and Terminal-Bench 4.0 at 53.90% against 56.06% — compares textGrain favourably with SynthID for text, and plans to open-source the technology. All vendor-reported, and the detector is closed to everyone outside the approved-researcher group at launch, so nobody has reproduced the numbers. Two practical consequences for a builder: the EU-only scope means identical text can carry a watermark in one region and not another, and the absence of a detected watermark proves nothing at all — OpenAI’s own wording is that short, edited, translated or unsupported output may simply fail to detect.

• Which model Free and Go users actually receive in Chat. OpenAI’s launch post says GPT-6 in ChatGPT is powered by “GPT-6 Luna for Free and Go tiers,” and the October 7 release notes repeat it: Free and Go use GPT-6 Luna. The help center article for GPT-6 in ChatGPT, updated the same day, still says those users receive GPT-5.6 Luna, still lists Free and Go as “Not included” at Medium, High, Extra High and Pro, and still repeats the GPT-5.6 Luna default in three separate sections. Both statements are the vendor’s, published within hours of each other on October 7, and they describe different models — one an API model at $0.10 and $0.50 per million tokens, the other its 5.6-generation predecessor. The October 8 free rollout did not reconcile them: the two vendor pages still describe the same tier differently, and no independent party has checked which model is actually answering.

• Whether the Pro configuration ever gets Intelligent UI, and what the answer would mean. At launch the capability is withheld from the level OpenAI says is powered by GPT-6 Astra, which is at least consistent with the Astra Pro configuration being about compute, effort ceiling or run length rather than feature access. If it stays withheld, “higher performance” has a concrete counterexample attached to it. If it arrives later, the interesting question becomes mechanical — whether the interface runs on Astra as well as on Sol, or whether Pro effort only ever sees interfaces composed by the smaller model — and OpenAI has published nothing either way. October 16 supplies the mirror image, and it is the more informative of the two facts: the parallel agents, the better tool use and the faster responses are documented for GPT-6 Pro and for no other Chat setting. If that asymmetry holds, the Pro setting’s difference from the cheaper levels is no longer only the undefined “higher performance” — it is a specific, named feature set that the Sol-powered levels may or may not receive later. OpenAI has published nothing about whether Instant through Extra High will get the parallel agents too.

The shortest honest summary after October 16 is this. GPT-6 Astra is real, priced, shipping, and generally available — Pro subscribers meet it in the chat picker as GPT-6 Pro, which OpenAI’s help center now says is powered by GPT-6 Astra, Plus accounts run GPT-6 Sol in Chat and Astra on smaller allowances through ChatGPT Work and Codex, and the Plus chat picker has no Astra entry at all. GPT-6-Astra-Pro is real as a plan-gating decision, its availability on OpenAI’s top plans is OpenAI-confirmed, and the most useful fact about it is still the one the company has not published: what, exactly, the higher performance is. September 29 moved that question twice without answering it; October 5 answered a different one — what the premium costs inside a plan — and October 7–8 answered a third, the free tier’s date, while attaching a counterexample to the capability reading and leaving the newest Sol out of Chat; and October 16 answered a fourth, the one the capability reading can actually claim — the Pro setting now has parallel agents, better plugin and tool use and faster responses that OpenAI documents for no other Chat setting, with no figure attached to any of them. The premium now has a name — Ultrafast, six times the standard token rate on the API and eight times the allowance burn inside a plan, and available on no Pro tier below the $500 one — and a product: dots, the always-on agents OpenAI says are powered by the same gpt-6-astra string, included with Pro and Business Premium but fenced off in the EEA, Switzerland and the UK at launch. On October 5 the company raised the floor under both: the subscription’s default speed for GPT-6 Astra and GPT-6.1 Sol is now roughly 50% faster by the vendor’s own account, a claim made by an executive on X, unmeasured from outside and attached to no price change. The same day it opened textGrain, its invisible text watermark, to opt-in API customers worldwide and committed to switching it on by default for eligible ChatGPT and Codex output in the European Union over the coming weeks — a provenance change that says nothing about which model you are calling. And on October 7 it put GPT-6 into ChatGPT’s default chat surface with Intelligent UI, dated to the paid tiers on October 7 and to Free and Go from October 8, on GPT-6 Sol rather than the Astra-powered Pro setting, at no rate-card change anywhere; the October 7 release notes also confirm the models behind Work and Codex were left alone, which is where GPT-6.1 Sol still lives. Around those have grown seven separate ways to buy the model, each with its own gate: a plan tier you buy, a speed flag you meter, a token price you meter, an application you are admitted to, an agent that arrives with the subscription and is fenced by market, age and workspace setting, a thinking level a plan is or is not allowed to select, and now a piece of behaviour the seat has and the token does not. None of them lets you simply buy a more capable Astra by the token, and the two you cannot meter at all are the agent and the parallel-agent harness it now runs inside. SimpleBench’s September 7 entry remains the only independent measurement of a difference between the configurations: 86.5%, a tenth of a point behind Claude Fable 5.1 and ahead of the 83.7% human baseline, against 83.6% for the standard model. If you build on an API key, gpt-6-astra is still the one model string that matters, and it is listed on OrcaRouter at OpenAI’s list price, which makes “does your routing platform carry it” a check you can run this afternoon rather than a hypothetical. Treat the absence of a measurement as the story: October 16 named three behaviours for the Pro configuration and quantified none of them, which leaves the same two open questions it started with — what the higher performance is, and where it is measured. The other half of October 16 is invisible to an API key by design: context reused across subagents, cached longer, routed for more cache hits, described only as more effective usage on a plan, with no figure and no line on any rate card — which makes it the second thing that day a token buyer cannot purchase. And if you were planning to buy your way into it, note which setting OpenAI is now investing in: it is the one the parallel agents run on, and the one Intelligent UI still skips.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.