OpenAI previews Ultrafast: Sol at 14x speed on Cerebras — a service tier, not a model, and still no price

OpenAI previewed Ultrafast today: GPT-5.6 Sol at up to 14x the speed, in the API, to a select group of customers. The important word is tier. This is not a new model — it is a new way to serve one that has been live since July 9, running on Cerebras hardware instead of OpenAI's standard stack.

What is confirmed, from OpenAI's own post. Ultrafast is a service tier that runs GPT-5.6 Sol up to 14x faster than Standard processing, generating up to 750 output tokens per second, launching first in the OpenAI API. It is in limited preview to a select group of customers, expanding as capacity grows. Early testers named are Jane Street, Podium, Basis and Rogo. Internally OpenAI reports using it for incident response — reading logs, traces and conversations while an outage is still unfolding — and for research workflows.

Cerebras published its own post the same day, and that is where the numbers live. Against output speeds reported by Artificial Analysis, Cerebras puts Sol on Ultrafast at 11x faster than Fable 5 and 5x faster than Opus 4.8 on Fast mode. On Humanity's Last Exam, all 2,500 questions took Sol Ultrafast 11 hours 11 minutes against 78 hours 27 minutes for Claude Fable 5 — described as comparable accuracy roughly 7x faster. On GDP-Val, a benchmark of economically valuable knowledge work, it reports a 5.6x end-to-end speedup with no quality degradation. Cerebras states its methodology: benchmarking run by Cerebras on July 31, 2026, using Sol and Sol Ultrafast at medium reasoning inside Codex.

Give them credit for dating and specifying that, because most vendor benchmarks do neither. Then read it for what it is: the 14x is OpenAI comparing its own tier to its own tier, and every cross-vendor figure is the hardware vendor benchmarking the product it sells against that product's competitors. Those can both be true and still not be independent.

The soft spot is the quality half. "Comparable accuracy" on HLE is asserted without scores — we are given two wall-clock times and an adjective. Speed is the easy thing to measure and the easy thing to publish; the claim that matters is that nothing was lost, and that claim arrives without a number attached. GDP-Val does say "no quality degradation", also without a figure.

There is also no price. A new service tier, a stated token rate, named enterprise testers, and nothing at all about what it costs. Tokens per second is not a purchasing decision without dollars per token — and dedicated inference hardware historically does not sell at commodity rates. Until a number appears, "14x the speed" is a capability, not an offer.

One thing we can check ourselves, because we did the work last week: the July 31 benchmark predates the Sol update OpenAI shipped on August 6. That update applies only to the chat experience in ChatGPT — OpenAI said so explicitly, and we verified at the time that no API id or price moved. So the Sol being measured on July 31 is the same Sol the API serves today. The dates look mismatched and are not.

A note on a leak, carefully. Six days ago a tracked account claimed Astra was "running at Cerebras-level speeds inside OpenAI." That phrase now has a concrete referent — OpenAI is in fact serving a frontier model on Cerebras. But the leak was about Astra, and this is Sol, so it is a near-miss rather than a hit, and Astra is the model OpenAI said on August 7 it is deliberately slowing down.

What this changed on our board. Nothing fired, and it could not have: the registry tracks models, and a serving tier of an already-live model resolves to the existing GPT-5.6 entry, whose hunt was disarmed the day it launched. That is a third distinct blind spot found today, after an unleaked point release with no entry at all and a catalog-only sweep that cannot see first-party launches. We have armed Ultrafast as its own entry — verified absent from all 411 OpenRouter and 6,297 models.dev ids first — so the moment it stops being a select-customer preview and becomes something you can buy, that fires. Which, for anyone waiting on a price, is the event worth waiting for.