Perplexity's Ultrafast Speed Setting: Astra and Sol 6.1 Get a Fast Lane Inside Computer

Perplexity announced on October 14, 2026 that GPT-6 Astra and GPT-6.1 Sol can now be run on an Ultrafast speed setting inside Perplexity Computer, for Max subscribers only. This is not a model launch and it should not be read as one: GPT-6 Astra shipped on September 3, 2026 and GPT-6.1 Sol shipped on September 29, 2026, and both have been reachable inside Computer for weeks. What changed in the last twenty-four hours is the picker. A Max subscriber choosing Custom in Computer can now select either model and set speed to Ultrafast, and Perplexity's stated reason is the one a subscriber would guess — both "generate tokens faster, completing long Computer tasks sooner." Three things Perplexity has not said matter more than the sentence it has, and the rest of this piece is about separating the two.

What Perplexity actually shipped

The announcement is a single post on Perplexity's own account, and it is short enough to quote in full: "Ultrafast versions of GPT-6 Astra and SOL 6.1 are now available in Perplexity Computer for Max subscribers. Both generate tokens faster, completing long Computer tasks sooner. Choose Astra or SOL 6.1 under Custom, then set speed to Ultrafast." That is the whole public record so far. Perplexity's changelog, checked on October 15, still ends at its October 5 entry, which covered Automations, inline visualizations and the arrival of GPT-6.1 Sol in the product — the Ultrafast setting has not been written up there yet. Any figure beyond the three sentences above, including what the setting costs in credits, is not in the announcement.

The mechanism the post describes does check out against Perplexity's own documentation, and that is worth noting because it tells you where in the product to look. Perplexity's help centre page on effort mode states that selecting Custom lets a user "choose an available model and its supported reasoning level, along with any speed settings offered for that model," and adds a line that answers a question the announcement leaves open: "There are no separate controls for supporting agents." In other words, Custom configures the orchestrator — the model Perplexity documents as deciding how to approach a task and delegating parts of it to supporting agents that can run on models from different providers. It does not reach into the sub-agents.

Perplexity also made the two models available on different timelines, which is the least confusing part of this. GPT-6 Astra has been in Computer for Pro and Max subscribers since September 5, 2026, when Perplexity framed the addition with its own WANDR evaluation — Astra at 0.682 and $11.98 per task, figures Perplexity published and nobody has independently reproduced. GPT-6.1 Sol arrived on October 5, 2026, where Perplexity's changelog says it is "available to eligible paid subscribers in Perplexity and Computer" and that it powers the Light effort setting. Ultrafast is the third layer on top of both.

The name is not a coincidence, and that is the open question

Ultrafast is not a word Perplexity invented. It is the name of the fastest service tier in O​penAI's own API, and the two models it covers in that API are exactly the two models Perplexity just added it to — no others. O​penAI's Ultrafast mode documentation reads: "Ultrafast mode is the fastest service tier in the O​penAI API. It is broadly available for GPT-6 Astra and GPT-6.1 Sol." The configuration is a request flag, not a distinct model: you set model to gpt-6-astra or gpt-6.1-sol and service_tier to ultrafast, and O​penAI schedules the same checkpoint on faster hardware.

Observe that O​penAI's own documentation names the tier "the fastest service tier in the O​penAI API," scopes it to exactly these two models, and configures it with a service_tier flag rather than a separate model id — all of which is visible on the page captured below.

Two things follow. The first is that Perplexity's Ultrafast may be O​penAI's Ultrafast — the names, the model pair, and the stated effect all line up — or it may be Perplexity's own faster serving of the same checkpoints, in which case the shared vocabulary is doing a lot of unearned work. Perplexity has not said which, and there is no way to tell from outside. The second is that the price implication runs in a specific direction. O​penAI bills its Ultrafast tier at six times Standard on every line of its rate card, and sells it on its own subscription surfaces — Codex and ChatGPT Work, on the $500 Pro plan and eligible Enterprise and Edu plans — so if Perplexity is reselling that tier, a $167-per-month Max plan is a materially cheaper door to it than O​penAI's own top consumer tier. If Perplexity is not reselling it, the comparison is meaningless. That distinction is the single most useful thing a prospective subscriber could learn, and it is the thing the announcement declines to answer.

A third detail from O​penAI's documentation is worth carrying over regardless of which answer turns out to be right, because it is the kind of thing that determines whether the setting is usable at all. Ultrafast for GPT-6.1 Sol supports US and EU data residency and global processing; Ultrafast for GPT-6 Astra supports US data residency and global processing only. If Perplexity's setting is the O​penAI tier, an EU-residency requirement rules Astra out of it. Perplexity has published nothing about residency for the setting, and a subscription product routes the call for you, so a subscriber may have no visibility into it at all.

What the speed actually buys, and where it stops

The honest description of what changes is narrow, and O​penAI's own phrasing is the most precise available: the tier "reduces the time between generated output tokens." That is a statement about how fast the model types, not about how long it thinks. Nothing about the weights changes — same checkpoint, same 1,050,000-token context window for both models, same answers.

For the workload Perplexity is targeting, that is genuinely useful. Computer is an agent: it plans, then takes a long series of steps, and each step ends in output that the next step waits for. If a task takes sixty turns, and every turn's output lands sooner, the saving compounds across the task, and the wall-clock win can be far larger than the per-token number suggests. That is the case Perplexity is making when it says long Computer tasks finish sooner, and for long multi-step tasks it is the right case to make.

Where it stops is the deliberating. A task whose wall-clock time is dominated by the orchestrator reasoning — weighing approaches, reading forty sources before writing anything — will not speed up by the amount the tier's marketing implies, because the slow part is the thinking, not the typing. Two more ceilings apply if the setting is the O​penAI tier. Ultrafast has separate rate limits from Standard and Fast mode, with Astra's defaults set lower than Sol's at every usage tier (500,000 tokens per minute at Build for Astra against 1,000,000 for Sol), so a heavy Computer task can run into a provisioning wall rather than a speed one. And O​penAI "strongly recommends" persistent WebSockets for agentic applications, warning that "without a persistent connection, network overhead can reduce the latency gains" — a recommendation aimed at developers that a subscriber cannot act on, because the connection is Perplexity's to manage and not the subscriber's.

O​penAI publishes one speed multiple and only one: GPT-6 Astra Ultrafast "generates tokens up to 8x faster than GPT-6 Astra in Standard mode in Codex." That is Astra's measurement, on O​penAI's own harness, not Sol's and not Perplexity's, and no equivalent figure exists for GPT-6.1 Sol. Perplexity's announcement carries no multiple at all. Treat the number you have seen quoted for this setting as Astra's, and as vendor-reported.

What Max costs, and the credit question nobody has answered

Max is Perplexity's top consumer tier at $167 per month billed annually, against $17 per month for Pro. What the extra buys on the model side is stated plainly on Perplexity's pricing page: frontier AI models, preferred model selection, priority access to new features, and "maximum Computer usage." The Ultrafast setting slots into that list as a fourth Max-only item, and it is consistent with how Perplexity has handled model access before — the models arrived for Pro and Max, the speed control arrives for Max.

Credits are where the arithmetic gets interesting, and where Perplexity has left a gap. Computer runs on credits, priced at 100 credits to the dollar, and the documented ranges are wide:

• Quick fact or definition — 15–70 credits, or roughly $0.15 to $0.70

• Multi-step research report — 350–950 credits, roughly $3.50 to $9.50

• Data dashboard with competitor collection — 2,400–9,800 credits, roughly $24 to $98

• Entire app or multi-day build — 15,000-plus credits, $150 and up

Max includes 10,000 credits a month plus a one-time 35,000-credit bonus that expires in about 30 days, and auto-refill can top the balance up. The documentation says credits "are based on the work Computer actually performs, not a fixed price per task," and that tasks requiring significant computing power use more. It does not say what a faster tier does to that number, and neither does the announcement. If Ultrafast is O​penAI's tier, the tokens it generates are billed to whoever holds the account at six times Standard — a plausible reason for Perplexity to stay quiet until it publishes a multiplier. Until it does, the only defensible advice is to watch the usage page on a first Ultrafast run before committing a long build to it, because the two credit ranges above are far enough apart that a surprise is expensive.

Two ways to buy the same speed, and who should care

There are now two doors to a fast lane on these two checkpoints and they are priced on completely different logic. Perplexity's is a subscription: a flat $167 per month, credits metered against work performed, the models and the connection managed for you, nothing to integrate. O​penAI's is metered: service_tier: "ultrafast" on your own account, no subscription required, and a published multiple of six times Standard on every line.

The metered path is the one you can price in advance, because the rate card is public. On short context, Standard GPT-6 Astra is $10.00 per million input tokens and $50.00 per million output; Ultrafast is $60.00 and $300.00. For GPT-6.1 Sol, Standard is $2.00 and $10.00 against $12.00 and $60.00 on Ultrafast, with cached input at $0.10 moving to $0.60. Long context doubles the input side and lifts output further, and the whole ladder is exactly six times Standard at every rung — there is no intermediate step between Fast at twice Standard and Ultrafast at six.

That arithmetic is worth doing before assuming the subscription is the cheap door. A heavy Ultrafast month on the API can outrun $167 quickly, and a light one will not come close; the subscription wins on predictability and the API wins on paying only for what you use. If you want to run the comparison yourself rather than reason about it, both checkpoints sit on one key at OrcaRouter as openai/gpt-6-astra and openai/gpt-6.1-sol at provider list price with zero markup, so the same trace can be priced on both models without opening a second account — and because list price is passed through unchanged, the meter matches the vendor's rate card rather than a marked-up version of it. The Ultrafast tier itself is a flag on your own O​penAI account and not a separately routable model, so it is not something a router can sell you; what a router can do is make the tier-vs-tier decision a one-line change instead of a migration.

Worth stating plainly, since the shape of this story invites the mistake: Perplexity Computer is a subscription agent product, not an API. There is no Perplexity endpoint that hands you GPT-6 Astra at Ultrafast speed for your own application. If what you want is a fast lane inside somebody else's agent, the October 14 setting is the thing to try, on a Max plan, with the usage page open. If what you want is a fast lane inside your own product, Perplexity's announcement does not reach you at all, and the O​penAI tier — or the same checkpoint served by whoever you already route through — is the relevant decision.

The three questions still open

Does Ultrafast cost more credits? Perplexity has not published a multiplier, and its credit documentation only says that consumption tracks the work performed. Given the models arrived on different plans and the speed control is Max-only, a price step of some kind is plausible; the announcement does not provide one, so the only honest answer today is that the first long task on the setting is an experiment with an unknown bill.

Is Perplexity's Ultrafast the same as O​penAI's Ultrafast tier? Unconfirmed. The names, the model pair and the described effect all match O​penAI's tier exactly, and the naming match is too precise to be an accident — but "an Ultrafast version of the model" is Perplexity's own wording, and a fast serving path of Perplexity's own would be described the same way. Residency behaviour would settle it for an EU customer, since O​penAI's tier for GPT-6 Astra is US-only and Sol's covers the EU. Perplexity has published nothing on residency here.

Do supporting agents get the fast lane? No — or at least not by this control. Perplexity's effort-mode documentation states that Custom configures the orchestrator's model, reasoning level and speed settings, and that there are no separate controls for supporting agents. So the delegation that happens beneath the orchestrator runs on models chosen by Perplexity, at whatever speed they run at. On a long task with a lot of sub-agent work, that is a real limit on how much of the win reaches you.

What to watch next

Two things would close most of this out, and both are Perplexity's to publish. A changelog entry — the product's own record is still a version behind — would confirm the setting is permanent rather than an experiment, and a credit figure would turn the purchase decision from a guess into arithmetic. The third thing, the O​penAI-tier question, may never be confirmed: it is not the kind of detail a subscription product volunteers, and from the subscriber's side the setting either makes a long task finish sooner or it does not. That outcome is testable today, on your own traces and your own usage page, which is more than can be said for most of what gets announced in a week like this one.

A product and pricing change is also the moment to check the record: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.