OpenAI's "o" Agent: What the DevDay 2026 Leak Actually Says, and What It Doesn't

The reported name is the least interesting part. What the leak says about the company's always-on agent is that it will be called "o", that it will sit inside every ChatGPT Pro tier, and that a Fast Mode — running on Cerebras silicon — is attached to it. The detail worth stopping on is smaller and stranger than that: the reference was reportedly spotted on the upgrade page for the $100 Pro plan, not the $200 one, in the same month the company stopped selling the $200 plan at all. There is no product page, no price, no model identifier and no company statement using the letter as a product name. The post behind this article is one account's summary of what it says it has seen, and the conference it hangs on is real: DevDay 2026 is September 29 at Fort Mason Center in San Francisco. The agent it competes with is real too — Gro​k Bot, the digital-coworker product SpaceXAI put into early beta on August 11. The model the agent would presumably run on is real and dated: GPT-6 Astra, generally available since September 3. Almost nothing else in the claim is.

What the leak actually consists of, sorted by weight

Three things are being reported, and they do not deserve to be read as one claim.

The screenshotted references are the strongest part. Reporting published September 26 says references to "O" appear across ChatGPT's own configuration — as a display name, and as a mail suffix written as "-o" — and that the same string was recently listed as a benefit on the upgrade page for the $100 ChatGPT Pro plan. Those are deployment-shaped artifacts rather than invention: configuration surfaces are where a company stages a product before it announces one, in the same way that feature flags are. They are also the weakest kind of artifact available, because a staged string can be renamed, repurposed or abandoned, and it carries no behaviour with it. The report is explicit about that gap: what the agent does, what it may access, whether it schedules anything, how long it remembers, and who gets it are all unstated.

The name is the second-order claim. A suspended @o account on X is cited as a possible reservation. A suspended handle is a data point about intent at some moment, not about what the product is, and the same letter has already been through OpenAI's naming department once — more on that below.

The third layer is the part that came from this week's signal rather than from reporting: that "o" appears in all Pro tiers, that Fast Mode is confirmed, that it is presumably bundled into a $500 subscription tier, and that Cerebras silicon is what makes the speed possible. "Presumably" is doing visible work in that sentence. The signal post itself is truncated in the capture the radar took — its fifth item begins "shared m…" and stops — so read the list as incomplete rather than as five settled points.

The naming problem nobody has checked

OpenAI has already used the letter. Its reasoning models shipped under o1, o3, o3-mini and o4-mini across 2024 and 2025, and the line has been wound down rather than expanded: o4-mini was pulled from ChatGPT in February 2026, GPT-4.5 left in June, and o3 was given a 90-day sunset announced in the model release notes on May 28 — a retirement that this blog covered at the time because users were reporting vanishing responses before the deadline.

So a bare "o" as the name of a consumer agent would collide with the company's own recent history, and would be ambiguous in the one place ambiguity is expensive: a model picker. That is not proof the leak is wrong — vendors recycle letters, and an internal codename often dies at the naming review — but it is a reason to treat the name as the most disposable part of the story. If "o" ships under a different label, every dated fact below still stands.

Why the timing is credible even though the product is not

The strongest evidence for a DevDay agent launch is not the leak. It is what OpenAI has already put in place over the four weeks before it.

• September 3 — GPT-6 Astra reached general availability as OpenAI's frontier model: 1.05M-token context, 128K maximum output, listed at $10.00 per million input tokens and $50.00 per million output, with cached input at $1.00 and cache writes at $12.50. Our own catalogue carries the same model under a 2026-09-04 listing date; Artificial Analysis dates the release to September 3.

• September 10 — OpenAI opened the Agents API in public beta, selling developers the managed Codex harness: OpenAI runs the orchestration, the context compaction and the recovery, and the developer supplies the task, the model, the tools and the environment. That is the infrastructure of an always-on agent, sold as a developer product.

• September 10 — the same day, new sign-ups and upgrades to the $200 ChatGPT Pro plan were paused, with demand for GPT-6 Astra given as the reason. The $100 tier, Plus, Go, Business and Enterprise and the API all stayed open.

• September 22 — GPT-6 Sol and GPT-6 Luna shipped at $2.00 / $10.00 and $0.10 / $0.50 per million tokens, which OpenAI has described as permanent rather than promotional, extending the cheap end of the line that the agent would draw on for routine work.

• September 26 — the Ultrafast serving tier's documentation expanded, with a speed selector offering Standard, Fast and Ultrafast spotted unpublished in the Responses API Playground and reporting pointing at a wider rollout around DevDay.

• September 29 — DevDay 2026, keynote livestreamed, Sam Altman opening. OpenAI's developer account posted "72 hours to OpenAI DevDay" on the same weekend the agent report landed.

Read those six lines together and the shape of the story is coherent: OpenAI productised the agent harness, sold it to developers, rationed the consumer seat that carries its best model, and has a keynote in four days. A consumer-facing version of that same harness — sold as a subscription rather than a per-token endpoint — is the obvious next move. None of that makes the name, the tier, or the Cerebras link verified. It makes the timing plausible, which is a different and weaker claim.

Fast Mode, Ultrafast and the Cerebras claim

The signal says Fast Mode is "confirmed" and that the Cerebras chip "takes effect". Both need separating from each other, because one is documented platform plumbing and the other is an inference about which chip serves which product.

Fast mode is a real, priced tier on OpenAI's API for the GPT-6 line: it bills at double the standard rate and buys roughly 2.5× output speed, and it is the same tier that was renamed from Priority processing on July 30, 2026 — the API accepts either service_tier: "fast" or service_tier: "priority". It is unavailable with EU data residency on at least one model in the line. Inside the ChatGPT subscription pool, reporting on the subscription credit structure puts Fast mode at 2.5× the credit rate rather than the API's 2×, which makes the fastest setting a slightly worse deal inside ChatGPT than outside it — a detail worth knowing before assuming a Fast-Mode benefit is worth what it sounds like.

Ultrafast is the Cerebras tier and it is a different animal. OpenAI announced it on August 13, 2026 as the first product of its compute partnership with Cerebras — a deal reported at roughly $10 billion over three years — and it delivers up to 14× standard throughput and up to 750 output tokens per second on GPT-5.6 Sol, the model it is currently restricted to. It is access-controlled, it has no published price, and as of this week it is still a waitlist tier rather than a general one. The Responses API reference lists ultrafast as a value for the service_tier parameter with the caveat that it is "currently available for gpt-5.6-sol".

Which is exactly why the leak's Cerebras sentence should be treated as a guess with a plausible basis rather than a confirmation. A consumer agent that runs continuously would be dominated by latency and cost per task, and Cerebras-class speed is a genuine answer to the first half of that. But there is no documented Ultrafast path for any GPT-6 model yet, no rate card for it, and nothing from OpenAI tying a consumer agent to that silicon. The direction of the inference is reasonable; the certainty in the phrasing is not earned.

What "available in all Pro tiers" would cost you

This is the part of the claim a reader can actually evaluate, because the tier structure underneath it is published.

• Monthly prices — the $100 and $200 ChatGPT Pro plans have both been on sale through September; new sign-ups to the $200 plan have been paused since September 10 and OpenAI has published no reopening threshold.

• The metered version of the same capability — GPT-6 Pro is capped at 200 messages a week on the $200 plan and 50 a week on the $100 plan, the latter drawn from an allowance shared with GPT-5.6 Sol Pro. Business tiers sit lower. Those are OpenAI's published figures and they are the real limit on what a "benefit on the upgrade page" is worth.

• Token arithmetic on the model underneath — at $10.00 / $50.00 per million, $500 of GPT-6 Astra API is about 50M input tokens or 10M output tokens, halving again in effective work above the 272K-input threshold where the whole request reprices at $20.00 / $2.00 / $75.00.

• The cheaper half of the line — the same $500 buys roughly 250M input or 50M output tokens on GPT-6 Sol at $2.00 / $10.00, and ten times that on GPT-6 Luna at $0.10 / $0.50.

If the agent arrives inside an existing subscription, the comparison stops being about list price and becomes about which seat you already pay for. If it arrives as its own line item, OpenAI has to justify it against a category where a rival has been shipping for six weeks. Either way, an always-on agent changes the shape of the bill: an assistant that keeps working while you sleep spends tokens without a human in the loop, and that is the workload a metered tier punishes and a seat absorbs. It is also the workload where cached input at one tenth of fresh input decides the total — a persistent agent re-sends an almost-unchanged context every turn, which is precisely the pattern a cache discount is built for.

That is the layer OrcaRouter operates at, and it is worth being plain about what it does not cover. We route OpenAI's models by the token with provider list price passed through at 0% markup, so a vendor rate change is live here the same day, with automatic failover across providers and a routing DSL for composing calls. GPT-6 Astra, GPT-5.6 Sol, GPT-6 Sol and GPT-6 Luna are all on one key. What we cannot sell you is a ChatGPT subscription or an agent that does not exist yet — there is nothing to route until OpenAI ships one.

The bar the agent has to clear is lower than it looks

Gro​k Bot is the comparison the leak itself makes, and the honest reading of it is that SpaceXAI has set a low bar. It went into early beta on August 11 in a $120-per-seat entry configuration bundled into Cursor Premium Teams, with individual tiers at $200 and $300 a month and usage beyond the included limits billing at token cost. Each bot gets its own cloud VM with a browser, a filesystem and a terminal; it signs into apps that have no API; it saves demonstrated workflows as routines and keeps going after the laptop closes.

What has not been published is the part that would settle anything. SpaceXAI has released no independent benchmarks or reliability data for it, the model router behind it is automatic and users cannot pin a model, early testers have been publicly unimpressed, and reviewers note the absence of an enterprise control layer — no scoped permissions, escalation rules or audit trail. Meta's Muse followed on September 8 into a category that is now three products wide.

That last row is the point of the whole comparison. There is no Artificial Analysis index for "did the agent finish the task without doing something catastrophic", so a persistent agent is bought on the same evidence a chatbot was bought on in 2023: demos and vibes. The one dated constraint nobody in this story mentions is OpenAI's own safety posture. GPT-6 Astra reached the company's Critical cybersecurity capability threshold, and OpenAI has warned that monitoring may slow, pause or stop legitimate tasks — explicitly including long-running jobs. An always-on agent is a long-running job by definition, in the platform that has already said it may interrupt one.

What would make this checkable, and when

Rumours about a product four days out are cheap, and the checks are specific.

• An OpenAI page, release note or help-centre entry that uses the name as a product name. Nothing has appeared so far, and the letter's history inside OpenAI makes this the test with the most room to fail.

• A configuration surface that survives a refresh — a genuinely purchasable plan, a listed allowance, a settings page — rather than a display string staged in a build.

• Any behaviour specification at all: what it can be granted, what it schedules, how long it remembers, whether it has its own mail identity in production or only in a config file.

• A price. "In all Pro tiers" is not a price, and a $500 tier does not become a fact because a separate report found $500 subscription strings in a different place.

• The Cerebras link either showing up in OpenAI's own serving documentation for the tier, the way Ultrafast's restriction to GPT-5.6 Sol did, or not showing up at all.

The decision this leaves a reader with is smaller than the leak implies, and it is not about the letter. If you are evaluating an always-on agent for work, the product that exists today is Gro​k Bot, and the sensible move is a short trial against your own workflow rather than a seat commitment — its $120 entry tier is a floor, not a ceiling, and usage beyond the included limits bills at token cost. If you are building, the Agents API has been in public beta since September 10 and is the part of this story you can actually test today, with the harness, the sandbox and the billing all documented.

And if you are pricing what any of this would cost to run, do it on the models rather than on the rumour. GPT-6 Astra at $10.00 / $50.00 per million tokens is live on OrcaRouter at OpenAI's own list price with no markup added, GPT-6 Sol and GPT-6 Luna are there at $2.00 / $10.00 and $0.10 / $0.50, and the cache read rate — $1.00 against $10.00 for Astra — is the line that decides what a persistent workload actually costs. An agent OpenAI has not announced cannot be benchmarked, routed or bought. The tokens underneath it can be, today, and the meter you build against them will still be the right meter if the name on the box turns out to be something other than a single letter.