OpenAI Ultrafast Mode: What 14x Means for Agents (2026)

OpenAI Ultrafast mode is an API service tier running GPT-5.6 Sol at up to 750 output tokens per second on Cerebras hardware, which OpenAI describes as up to 14x Standard processing. Previewed August 13, 2026, it has no published price or availability date.

TL;DR

Ultrafast is a speed tier, not a new model. The same GPT-5.6 Sol weights, different silicon underneath. OpenAI announced it on August 13, 2026, with the headline figures of up to 750 output tokens per second and up to 14x Standard processing.¹

The parts the headline number leaves out:

• The only end-to-end figure anyone has published is 5.6x. On GDPval, running inside Codex, Cerebras reported "a 5.6x end-to-end speedup with no quality degradation."² The two numbers are not rivals — 14x is an up-to ceiling on some unstated part of a request, 5.6x is a measured result on one benchmark — but 5.6x is the closer analogue to a real workload.
• This is not OpenAI's first speed tier. Priority processing was renamed Fast mode on July 30, 2026, and already offers up to 2.5x Standard for gpt-5.6-sol.³ Ultrafast is the second rung above Standard, not the first.
• Ultrafast has no published price. As of August 17, 2026, OpenAI's pricing page lists gpt-5.6-sol under Standard, Batch, Flex and Fast mode only — there is no Ultrafast row.⁴
• The nearest available anchor, and a weak one: OpenAI and Anthropic both charge exactly 2x standard rates on the tiers they each describe as "up to 2.5x". GPT-5.6 Sol goes from $5/$30 to $10/$60 per million tokens; Claude Opus 5 and Opus 4.8 from $5/$25 to $10/$50.⁴⁵ OpenAI's own multiplier is not constant, though — it ranges from 1.75x to 2.5x across its catalog.
• The premium may fall on the wrong tokens. Anthropic documents that its tier speeds up output rather than time to first token; OpenAI has not said whether Ultrafast speeds up anything beyond generation. In the tiers that are priced today, the premium applies to input as well as output — and a turn of 50,000 input and 500 output tokens is 94% input cost at published rates.
• Access is limited preview, select customers, with a notification form for everyone else.¹

What You'll Learn

• What Ultrafast actually is, and which parts of the announcement are measured versus marketed
• Why the 14x and 5.6x figures are both honest, what each one actually measures, and which is the better guide for your agent
• Which part of a request a speed tier actually accelerates, and why speeding up generation tends to leave prompt processing deciding your wall clock
• Where Ultrafast sits on OpenAI's service-tier ladder, what the priced rungs cost today, and why the premium may fall on the wrong tokens for agents
• What mixing speed tiers inside one agent may do to your prompt cache
• What you can and cannot do with Ultrafast today

What Is OpenAI Ultrafast Mode?

Ultrafast is, in OpenAI's own words, "a new service tier that runs GPT‑5.6 Sol up to 14× faster than Standard processing, launching first in the OpenAI API."¹ It is powered by Cerebras and "generates up to 750 output tokens per second".¹

The strategic claim underneath is more interesting than the number. OpenAI frames it as breaking a tradeoff: "Until now, getting real-time speed typically meant choosing a smaller or more specialized model."¹ That reads partly as a reference to its own catalogue — in February OpenAI released GPT-5.3-Codex-Spark as a research preview, "a smaller version of GPT-5.3-Codex, and our first model designed for real-time coding," also running on Cerebras.⁶ Ultrafast is the claim that you no longer have to drop down to one. The company's summary phrase for the tier is more useful work per second.¹

Cerebras explains the mechanism as a memory-bandwidth story. "Fast frontier inference is a data movement problem," its engineering post argues: on GPUs, large-model inference is bottlenecked by shuttling weights between on-chip and off-chip memory for every token. Cerebras packs "44 GB of SRAM on each wafer-sized chip," so "weights stay on-chip, and tokens flow uninterrupted through model layers pipelined across wafers."²

OpenAI named five scenarios: incident response and reliability, financial research and security, customer support and voice, commerce, and live research and experimentation.¹ The support example is the most explicit about the work being multi-step — resolving issues in real time "even when finding the answer requires multiple steps or systems" — but the research scenario describes an iterative loop too, and none of the five is a single-shot question.¹

One detail matters if you are quoting the announcement in an agent context: OpenAI's page never uses the word "agent" at all. It says only that early testing spans "coding, commerce, financial research, support, and other interactive applications," and describes the workflows functionally. The explicit agent framing belongs to Cerebras, which writes that "with Ultrafast, you can now put agents on the critical path of problems where every second counts."²

The 14x Reads Like a Generation Number. The Measured Number Is Smaller

Here is the figure OpenAI's own announcement leaves out. In the same set of announcements, Cerebras reported that on GDPval — OpenAI's benchmark of real-world economically valuable tasks across 44 occupations, whose deliverables "span documents, slides, diagrams, spreadsheets, and multimedia"⁷ — "Ultrafast delivered a 5.6x end-to-end speedup with no quality degradation."² That test was run by Cerebras on July 31, 2026, comparing GPT-5.6 Sol against GPT-5.6 Sol Ultrafast "on medium reasoning within Codex."²

Both numbers are honest. They measure different things. The 14x is a ceiling; OpenAI describes it as "up to 14× faster than Standard processing" and never says which part of a request it is measuring.¹ The 5.6x is what Cerebras recorded running a real harness over a real benchmark — closer to what a workload experiences, though as the caveats below show, not a clean read on an agent loop either.

The gap is Amdahl's law doing its usual work. A step in an agent loop spends its wall clock on several things: the request round trip, prompt processing before the first token appears, generating the response, and then executing whatever tool the model asked for. A throughput tier targets one of those. Treat the step as the sum of its parts — an approximation a well-built harness can partly beat, as the next section explains — and the arithmetic is standard: if generation is the fraction p of a step's wall clock and it runs s times faster, overall speedup is 1 / ((1 − p) + p/s).

Three caveats before running that backwards. Cerebras does not say whether its "end-to-end" figure is per-task latency or aggregate throughput across a concurrent benchmark run, and single-step Amdahl's law does not apply to the latter. Because 14 is an up-to figure it is a ceiling on s, not a measurement of it. And using it as the generation multiplier assumes the 14x is a generation number at all — a reading the 750-tokens-per-second figure supports but OpenAI has never confirmed.

With those stipulated: set s to 14, and Cerebras's 5.6x implies roughly 88% of the GDPval wall clock was token generation. That is a very generation-heavy mix, which is what you would expect from a benchmark whose output is finished work products. Because 14 is a ceiling on s, 88% is a floor on p rather than an estimate — if the sustained speedup on that run was 8x rather than 14x, the same 5.6x implies p ≈ 94%.

The same asymmetry runs through the table below, whose arithmetic rows all assume you receive the full ceiling multiplier on the generation phase:

Amdahl's law applied to the 14x figure, assuming generation is accelerated uniformly at the ceiling and every other term is unchanged. The 5.6x row is Cerebras's published measurement, and the percentage beside it is solved from it; every other row is arithmetic, and each is a best case rather than an expectation.

One number worth extracting for your own use: 750 tokens per second at 14x implies a Standard-tier baseline of about 54 output tokens per second. OpenAI does not publish a Standard-tier throughput figure, and the derivation only holds if both headline numbers are realized together. For scale, the only throughput figure OpenAI commits to as a service level for this model is a Fast mode target of more than 80 tokens per second, and that applies to Enterprise customers.⁸ But it is a useful sanity check — if your own measured Standard throughput for gpt-5.6-sol is comfortably above 54 tokens per second, the multiple available to you is correspondingly below 14.

One more reason to treat 5.6x as an upper marker rather than a forecast: GDPval is, in OpenAI's own framing, a one-shot evaluation — a task in, a deliverable out.⁷ It is not a multi-turn loop, so it never pays a cost that agents pay constantly: resending and reprocessing a growing context, turn after turn. An agent whose steps are long reasoning traces with occasional cheap tool calls sits near the top of the table. An agent that fires a dozen short tool calls, each waiting on a slow database or a third-party API, sits near the bottom, where the model was never the bottleneck and a faster decoder has little left to win.

Cerebras's other headline is more spectacular and less relevant. Running Humanity's Last Exam — a 2,500-question benchmark built by the Center for AI Safety and Scale AI⁹ — Sol Ultrafast finished in 11 hours 11 minutes against Claude Fable 5's 78 hours 27 minutes, which Cerebras describes as "comparable accuracy nearly 7× faster."² The exact ratio is 7.01, so if anything the phrasing undersells it. But this is a cross-model, cross-harness, cross-date comparison, not a tier-to-tier one: Cerebras states the two runs used different harnesses on different days — Codex on July 10 for Sol Ultrafast, Claude Code on July 13–15 for Fable 5, both at the reasoning setting each vendor labels xhigh.² The stated effort level was the same on both runs, but two different models still emit different quantities of reasoning tokens, which could account for much of the gap without the hardware contributing anything. It tells you Sol Ultrafast is fast. It is not evidence for how much faster Ultrafast is than Standard.

All of these are vendor-run numbers. Cerebras's own site footer says so: "Performance comparisons are based on third-party benchmarking or internal testing. Observed inference speed improvements versus GPU-based systems may vary depending on workload, configuration, date and models being tested."²

Which Latency Term a Speed Tier Moves, and Why Nobody Will Tell You

Speed tiers do not accelerate a request uniformly, and which part they accelerate decides whether Ultrafast is worth anything to your agent. Vendors that sell one do not even agree on the answer.

Anthropic ships its own speed tier — confusingly, also called Fast mode — and its documentation is the closest published analogue, unusually candid about mechanics. It states that "speed benefits are focused on output tokens per second (OTPS), not time to first token (TTFT)," and repeats the point in its considerations list.¹⁰ Time to first token is mostly the cost of processing your prompt, plus queueing and the network round trip. Faster decoding makes the model emit faster; on Anthropic's platform it is explicitly not a promise that it starts faster.

xAI documents the opposite. Its tier is called Priority Processing — a name OpenAI used for Fast mode until July, and one Google uses too, for three different products. xAI's bills at "a premium per-token rate" and claims gains on both ends: "lower time-to-first-token (TTFT) and faster inter-token latency (ITL), especially during periods of high demand."¹¹ That is a different mechanism — queue position rather than faster silicon. Google sells a third variant, a Priority tier that "routes requests to high-criticality compute queues" at 75–100% above standard rates, with overflow traffic gracefully downgraded rather than rejected.¹²

So a premium tier can buy you a better place in the queue, a faster decoder, or both, and the vendors selling one do not describe the same product. That matters when you are reasoning about which part of your agent's wall clock will actually shrink.

OpenAI sits in between. Its Fast mode page states that tier's latency commitment as a floor on tokens per second rather than as a time to first token, which points Anthropic's way without saying so.⁸ For Ultrafast it has said only that the tier "generates up to 750 output tokens per second"¹ — a statement about generation that leaves prompt processing unaddressed. Cerebras's mechanism does not settle it either: keeping weights resident in on-chip SRAM might speed up prompt processing as well as generation, or might not. Nobody outside the preview can say, and for anyone running agents it is the question that matters most.

The distinction matters more for agents than for chat, and in the uncomfortable direction. A chat user waits once, at the start. An agent loop resends a growing transcript on every single turn: system prompt, tool definitions, prior messages, and every tool result it has accumulated — file contents, search results, page dumps. That prompt grows unless you actively compact it, and absent a cache hit, processing it is a cost paid again on every turn.

Then Amdahl's law bites from the other side. Suppose a step is one second of prompt processing plus thirty-five seconds of generation. Prompt processing is under 3% of the step, easy to ignore. Accelerate generation 14x and leave prompt processing untouched, and the step becomes one second plus two and a half — the term you were ignoring is now 29% of the wall clock. To whatever extent a tier accelerates generation and not prompt processing, it makes prompt processing the thing that matters. The faster the decode, the more of your remaining latency lives in the parts it does not touch — which is exactly why OpenAI's silence on the question is worth noticing.

Whether that matters to a specific agent comes down to one ratio: context length to output length, per step. The decode-bound end of that spectrum has a great deal to gain from a faster decoder; the prompt-bound end has almost nothing to gain. An agent that accumulates context over a long run travels from the first toward the second as it goes, which means the tier gets less useful the longer the agent runs.

One simplification is worth retiring: that an agent step is fully serial because nothing can happen until the model finishes. That is not quite true, and it is why the Amdahl model above is an approximation rather than a law. Tool calls stream incrementally. Anthropic's documented stream opens a tool_use block carrying the tool's name before any of its arguments arrive, then delivers the arguments as input_json_delta fragments, with an opt-in fine-grained mode for lower latency still.¹³ Nothing in the protocol stops a harness from acting on the name — warming a connection, acquiring a token, booting a sandbox — while the arguments are still being generated. The gains are bounded, since Anthropic notes that current models emit one complete key and value at a time, so there may be pauses between events,¹³ and non-idempotent tools still have to wait for complete arguments. But "every step blocks on the full response" describes a naive harness, not a necessary one.

OpenAI described the underlying shape itself when it announced the Cerebras partnership back in January: "When you ask a hard question, generate code, create an image, or run an AI agent, there is a loop happening behind the scenes: you send a request, the model thinks, and it sends something back."¹⁴ Ultrafast compresses the thinking. The round trip, the prompt processing and the tool are a different matter — and unless all of them happen to shrink by the same factor, an end-to-end figure will land below the headline. That is the most likely shape of the gap between 14x and 5.6x, though neither company has explained it. If you want a sense of how quickly those loops accumulate spend as well as time, our breakdown of AI agent token costs per task covers the cost side of the same arithmetic.

Where Ultrafast Sits on OpenAI's Tier Ladder

Ultrafast is easy to read as OpenAI's speed tier. It is the second one. OpenAI's docs record that "Priority processing was renamed Fast mode on July 30, 2026," and that it "increased the speed at which Fast mode operates for gpt-5.6-sol to make it up to 2.5× faster than Standard processing."³ Developers opt in with service_tier: "fast", or with "priority", which the docs say "provides the same behavior for supported models" — though on GPT-5.6 and earlier the response reports priority whichever of the two you sent.³ A project-level default is also available.³

Here is the ladder for gpt-5.6-sol at short context, per OpenAI's pricing page as of August 17, 2026:⁴

Three things fall out of that table.

First, Fast mode costs exactly 2.00x Standard for gpt-5.6-sol — on input, cached input and output alike, and the same holds for its Terra and Luna siblings.⁴ Anthropic landed on the same ratio: Claude Opus 5 and Opus 4.8 cost $5/$25 per million tokens at standard rates and $10/$50 in Fast mode,⁵ for "up to 2.5x higher output tokens per second."¹⁰ Two vendors, both charging double. The 2.5x figures are not strictly comparable — Anthropic's is explicitly output tokens per second, OpenAI's says only "faster than Standard processing"³ — but the price multiple is identical either way.

Treat that as the nearest available reference point rather than a formula. A doubling is also simply the most obvious number in pricing. And OpenAI's own premium is not a constant even within its own catalog — the same pricing page charges 2.5x for gpt-5.5 in Fast mode ($5/$30 to $12.50/$75), 1.8x for gpt-5-mini, and 1.75x for gpt-4.1.⁴ Ultrafast is also, in OpenAI's words, "powered by Cerebras"¹ rather than by whatever serves Standard, which plausibly gives it a different cost structure — though OpenAI has not described the deployment topology. And there is no reason a speed premium has to be linear, per-token, or even publicly listed at all.

Second, and more actionable: the premium may fall on the wrong tokens for agent work. As the previous section covered, these tiers are documented — where they are documented at all — as accelerating output. What both priced tiers charge for is input as well as output. Agent turns tend to be input-heavy, because the whole transcript goes back up the wire each time — and the further that skews, the worse the trade gets. Take a turn of 50,000 input tokens and 500 output tokens: at published rates it costs $0.265 on Standard, of which $0.25, or 94%, is input.⁴ Moving that turn to Fast mode takes it to $0.53, and on Anthropic's documented mechanic the extra cost is buying speed on the 500 output tokens rather than on the 50,000 input ones. Prompt caching cuts the absolute numbers hard, but not the premium: the cached-input rate doubles alongside everything else, so a cached turn still costs twice as much in Fast mode as in Standard. If Ultrafast is priced in the same shape, the arithmetic gets worse as contexts grow, which is precisely the direction agents move.

Third, the gap in that table is the real story for developers. There is no Ultrafast row on the pricing page, no Ultrafast guide among the API docs, and no documented service_tier value for it. The only trace in the API documentation is a changelog line announcing the preview.¹⁶ The one commercial detail OpenAI has published sits on its Reserved Tier page, which notes that Reserved Tier "is also available for Ultrafast mode, which is currently in limited preview"¹⁷ — so the capacity will be pre-purchasable, even though the per-token rate is not public. Nor has OpenAI said whether Ultrafast supports prompt caching, the full context window, structured outputs or parallel tool calls — any one of which, if absent, would matter more to a long-running agent than the speed does. Those are worth asking about rather than assuming, because Cerebras's stated advantage comes from what it keeps in on-chip memory,² and a long agent context is exactly the kind of state that has to live somewhere. None of it is answerable from outside the preview.¹

The Prompt-Cache Risk in Mixing Speed Tiers

The obvious optimization, once a very fast tier exists, is to mix: run the latency-critical steps of an agent on the fast tier and the rest on Standard. There is a documented reason to be careful.

Anthropic spells it out: "Switching between fast and standard speed invalidates the prompt cache. Requests at different speeds do not share cached prefixes."¹⁰ For an agent, that is expensive twice over. Agent loops resend a growing conversation on every turn, and prompt caching is what keeps that affordable — a cache read costs a tenth of base input on Anthropic's platform, and a tenth on OpenAI's Standard tier too ($0.50 against $5.00).⁴⁵ An agent alternating tiers turn by turn does not merely lose the discount; if its harness keeps setting cache breakpoints it pays Anthropic's cache-write rate as well — 1.25x base input for a five-minute write, 2x for an hour — landing above full input price while believing it is saving money.⁵

The second cost is the one the previous section was about. A blown prefix cache means the whole context gets processed again, and on a large agent context that is real seconds added to every turn — a latency regression that can plausibly exceed whatever a faster decoder was buying.

OpenAI has not published equivalent guidance. Its Fast mode documentation confirms that "cached input discounts still apply to Fast mode requests" and its pricing page lists a separate cached-input rate for the Fast tier, but neither states whether a prefix cached at one tier is readable from another.³⁴ For Ultrafast specifically there is a structural reason to be pessimistic: OpenAI has described it as a distinct tier running on Cerebras hardware,¹ and cached state does not obviously travel between separate deployments. Design as though tiers do not share caches, and ask before assuming otherwise.

Fast mode carries other operational sharp edges that a faster tier on the same platform will plausibly share. OpenAI applies a ramp rate limit: if traffic ramps too aggressively, the system "may downgrade some Fast mode requests to standard speeds and charge standard rates," with the response reporting service_tier: "default" — which means you must read the returned tier rather than assume you got the one you asked for. The limit "may apply" when sending at least 1 million tokens per minute and increasing that by more than 50% within 15 minutes.³ The docs also advise avoiding large ETL or batch jobs in Fast mode.³ Anthropic's version has a dedicated rate limit separate from standard Opus limits, returning 429s with a retry-after header when exhausted.¹⁰ Both are worth designing around. OpenAI notes that Standard and Fast share a model's rate limit,³ so generating several times faster brings a per-minute ceiling closer rather than raising it — and time spent in backoff is wall clock like any other. Speed tiers are capacity-constrained products, and they need a Standard fallback path behind them.

The Hardware Bet Behind It

Ultrafast is the second product to come out of a deal announced on January 14, 2026, when OpenAI said it was "partnering with Cerebras to add 750MW of ultra low-latency AI compute to our platform."¹⁴ The first was Codex-Spark, which OpenAI called "the first milestone in our partnership with Cerebras."⁶ Cerebras described it as a multi-year agreement to deploy 750 megawatts of wafer-scale systems, "making it the largest high-speed AI inference deployment in the world" — its characterization, not an independently verified one.¹⁸ OpenAI said at the time that it would "integrate this low-latency capacity into our inference stack in phases," with the capacity coming online "in multiple tranches through 2028."¹⁴

That timeline explains the shape of the launch. A limited preview gated on capacity is what you would expect seven months into a rollout that still has two years to run. Andrew Feldman, Cerebras co-founder and CEO, framed the thesis in January: "Just as broadband transformed the internet, real-time inference will transform AI, enabling entirely new ways to build and interact with AI models."¹⁴ Cerebras's own homepage markets Sol Ultrafast as "The World's Fastest Frontier Model" — again, its own claim, and not one made in the body of its technical post.¹⁹ For background on the company behind the silicon, see our earlier coverage of Cerebras's wafer-scale chip and Nasdaq debut.

The customers OpenAI quoted are consistent with the latency-sensitive framing: Jane Street, Podium, Basis and Rogo.¹ The most useful quote is Basis co-founder Mitch Troyanovsky's, because it identifies the actual constraint: "Oftentimes the barrier to truly fast products is not just tokens per second, but also model intelligence, and ultrafast combines both."¹

The Same Day, Google Made a Different Bet

The same day, Google announced Gemini 3.7 Flash, calling it "our most intelligent workhorse model yet for coding and agents."²⁰ The pitch runs on a different axis entirely. Google's announcement publishes no latency or throughput figure for the model; it is an intelligence-per-dollar argument, citing benchmark gains over Gemini 3.6 Flash at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens, rising to $1.50 and $7.50 on January 1, 2027.²⁰ Google does sell latency — Gemini 3.7 Flash is one of the models eligible for its Priority tier¹² — but it sells it separately, as a queue upgrade, rather than building the announcement around it.

Two routes to the same destination, announced the same day. OpenAI is selling frontier intelligence made fast; Google is selling cheaper intelligence made better. What is worth noting is not who wins but what both are aiming at: Google names coding and agents explicitly, and OpenAI's early-customer list is dominated by interactive, multi-step work even though its post never uses the word. A third route shows up in the model routing we covered on August 14 with NVIDIA's NeMo Switchyard agent model router, which splits the difference by sending each agent step to a different model entirely.

What You Can Actually Do Today

Very little, and that is the honest headline. Ultrafast is in limited preview with a select group of customers, expanding "as capacity grows," and what is on offer publicly is a form to be notified when access expands.¹ There is no price, no GA date, and no documented API parameter.¹³⁴

What is worth doing now is measuring, so you can evaluate the tier the day it opens up. Instrument each agent step into four buckets — network round trip, prompt processing, token generation, and tool execution — and record token counts and the p95 alongside the mean. The generation share tells you which row of the table above you occupy and therefore the ceiling on what any decode tier can buy you; the prompt-processing share tells you what will dominate once you get it; and the p95 tells you what your users actually experience, which is rarely the average.

Then ask the question the announcement carefully avoids: what is a second worth here? In the scenarios OpenAI leads with — an unfolding incident, a live customer call, a shopper still deciding¹ — seconds carry a dollar figure, and a fast tier can justify a steep premium. For agent work that runs unattended, like nightly refactors, review bots and bulk document processing, a second is worth close to nothing, and the correct move is the opposite direction: Flex and Batch are both half the price of Standard.⁴ Between those poles sit the interventions that cost no premium at all — parallelizing tool calls, lowering reasoning effort, compacting context, caching tool results, cutting steps out of the loop — several of which trade something for the speed, but none of which require a new tier.

One migration cost worth pricing in: "same model, different silicon" is not automatically "identical outputs." Cerebras reports "no quality degradation" on its own benchmarks,² and Anthropic says of its analogous tier that there is "no change to intelligence or capabilities."¹⁰ Neither is a guarantee about your pinned evals and hand-tuned prompts. Budget for re-validation.

None of which should be surprising, given how recently any of this became buyable at all. GPT-5.6 Sol itself was previewed in June 2026 under government-gated access restrictions before opening up more broadly, and Anthropic's competing Fast mode is still labeled a research preview requiring an account manager or a waitlist.¹⁰ Frontier speed, at both labs, is currently something you apply for.

Sources