Last updated: August 14, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
On August 13, 2026, OpenAI and Cerebras announced Ultrafast, a new API service tier that runs GPT-5.6 Sol up to 14x faster than standard processing. It's not a toggle you flip in ChatGPT settings yet — it's a limited preview, and getting in takes a specific application, not a subscription upgrade.
Short answer: Ultrafast is a preview API tier that runs GPT-5.6 Sol on Cerebras wafer-scale hardware at up to 750 output tokens per second, about 14x faster than Standard. It's currently limited to a hand-picked group of enterprise customers who apply through OpenAI's signup form. No public price, rate limit, or general-availability date has been announced yet.
In my testing of GPT-5.6 Sol on the standard API tier since it went broadly available in July, mostly for coding-agent and research workflows, the promise of Ultrafast is obvious the moment you've watched a long agentic run crawl through fifteen tool calls in real time. OpenAI hasn't opened Ultrafast to me directly — access is still restricted to a short list of enterprise partners while the two companies watch how the extra speed changes real products — so this guide is built on OpenAI’s own announcement and Cerebras’s technical writeup, plus what I ran myself on GPT-5.6 Sol's standard tier. I'll update it with hands-on Ultrafast numbers the moment that access opens up.
What you'll need
You need an OpenAI API account already using (or ready to use) GPT-5.6 Sol — Ultrafast is a service tier on top of that model, not a separate product. You'll also need a concrete workload to describe: what task you're running, how many requests per minute you expect, and why standard latency is actually a problem for you, not just an annoyance. OpenAI's signup form asks for workload details, latency requirements, and expected usage volume, so vague answers slow down review. Budget for the fact that GPT-5.6 Sol itself bills at $5 per million input tokens and $30 per million output tokens on the standard tier — Ultrafast hasn't gotten its own published rate yet, so plan around the existing Sol price until OpenAI says otherwise.
Step-by-step: Accelerating GPT-5.6 Sol Ultrafast
1. Confirm GPT-5.6 Sol is actually the bottleneck
Before you apply, check whether your latency problem is model speed or something else — network round-trips, a slow downstream API, or an oversized prompt. Ultrafast only helps the token-generation part of the pipeline; if your delay is elsewhere, 14x faster output won't fix it.
2. Get your workload numbers in order
Pull real numbers from your logs: average output tokens per response, requests per minute at peak, and the latency your users actually notice versus the latency you're guessing at. OpenAI is prioritizing applicants with production use cases — voice, live customer support, commerce, developer agents, financial research, and security response are the categories it's called out by name.
3. Apply through the Ultrafast signup form
Submit the request through OpenAI's official form and describe the workload from step 2 in specific terms. Skip generic language like "we want faster responses" — reviewers are triaging a small preview cohort, and a request tied to a named product and a real latency target reads very differently from a general interest signup.
4. Test on a low-stakes endpoint first
If you're accepted, don't point Ultrafast at your main production traffic on day one. Run it against a staging environment or a low-priority feature first, and compare output quality against your existing Standard-tier results on the same prompts. OpenAI reports no quality loss on the GDP-Val benchmark, but "no loss on a benchmark" and "no loss on your specific prompts" aren't guaranteed to be the same thing until you check.
5. Watch cost before you migrate anything real
Track token spend closely during the trial. Faster generation doesn't reduce token counts — output tokens still bill the same way — and until OpenAI publishes a dedicated Ultrafast rate, a spike in request volume during testing could cost more than you planned. My guide to managing AI coding costs at scale covers the tracking habits worth setting up before any high-throughput rollout.
GPT-5.6 Sol: Standard vs. Ultrafast
| Standard tier | Ultrafast (preview) | |
|---|---|---|
| Access | Open to all API accounts | Application-only, select customers |
| Output speed | Baseline | Up to 750 tokens/sec, ~14x faster |
| Hardware | OpenAI's default inference stack | Cerebras wafer-scale engine |
| Pricing | $5 / $30 per 1M tokens (in/out) | Not yet published |
| Context window | 1M tokens | 1M tokens (same model) |
| Best for | General use, cost-sensitive workloads | Voice, live support, incident response, agents |
Example prompts and requests you can copy
These are the kinds of requests worth using once you're testing Ultrafast, since they stress the exact things a speed tier is supposed to fix — turn-taking latency and long agentic chains:
- "Summarize this incident log in three bullet points and flag anything that needs escalation right now." (tests time-to-first-useful-output for on-call workflows)
- "Draft a customer support reply to this ticket in a friendly, concise tone, under 80 words." (tests conversational round-trip speed)
- "Given this error trace, list the three most likely root causes in order of probability." (tests reasoning latency under pressure)
- "Walk through this multi-step coding task one tool call at a time, and tell me what you're doing before each step." (tests agentic loop speed, where standard latency compounds fastest)
- "Read this transcript and generate a spoken-style reply I could use in a voice assistant, under 40 words." (tests the voice use case OpenAI specifically named)
Run each one on Standard first and time it yourself, so you have a real before number instead of guessing at the improvement once Ultrafast is live for you.
Common mistakes to avoid
The biggest one: assuming there's a self-serve toggle somewhere in the API dashboard. There isn't yet — Ultrafast is invite-based, and looking for a settings switch just wastes time. Second, applying with a vague pitch instead of real numbers; "we want it faster" gets deprioritized behind applicants who list actual requests-per-minute and a named product. Third, budgeting as if Ultrafast will cost the same as Standard — it might, but OpenAI hasn't confirmed that, and assuming it does before launch can throw off a cost forecast. Fourth, skipping the quality comparison step because OpenAI's own benchmark says there's no degradation — that's their number on their benchmark, not yours on your prompts, so check it yourself once you have access. Fifth, migrating all production traffic at once instead of a slow rollout; a preview tier is exactly the kind of thing that can hit capacity limits without warning.
Tools that make this easier
If you're building the kind of agentic workflow Ultrafast is aimed at, my guide to how to use ChatGPT Codex covers the coding-agent side of GPT-5.6 Sol, and how to use ChatGPT agent mode walks through the consumer-facing agent controls that share the same latency pain points. If speed is the whole reason you're reading this, it's worth knowing GPT-5.6 Sol isn't the only fast option — DeepSeek V4 Flash targets the same low-latency niche from a different price point, and my GPT-5.6 vs. Claude Fable 5 comparison covers how Sol stacks up against Anthropic's fastest model on tasks where speed matters as much as accuracy. And before you commit engineering time to an Ultrafast migration, AI is getting way too expensive is a useful gut-check on why token spend deserves the same scrutiny as latency.
My take on Ultrafast, for now
The speed numbers are real and independently reported, not just a vendor claim sitting alone on a landing page — outside coverage confirmed the 750 tokens/second and 14x figures directly from OpenAI's briefing. What's missing is everything that turns a preview into something you can plan around: a published price, a rate limit, and a timeline for wider access. If your product genuinely needs sub-second responses from a frontier model — voice, live support, security response — apply now and get in line. If you're just curious, it's worth reading about, but there's nothing to build on top of yet. My full guide to how to use ChatGPT is a better place to start if you're newer to GPT-5.6 Sol generally, before you worry about which service tier it runs on.
Frequently Asked Questions
Is GPT-5.6 Sol Ultrafast free?
No. It runs on top of GPT-5.6 Sol's existing API pricing of $5 per million input tokens and $30 per million output tokens, and OpenAI hasn't announced whether Ultrafast adds a premium on top of that rate.
How long does it take to get access to GPT-5.6 Sol Ultrafast?
There's no published timeline. OpenAI describes access as expanding "as capacity grows," and approvals are going to applicants with the clearest production use case first, not in the order applications arrive.
What is the easiest way to test Ultrafast once I'm approved?
Point it at a low-stakes staging endpoint first, run the same prompts you already use on Standard, and compare both speed and output quality side by side before touching production traffic.
Does Ultrafast change GPT-5.6 Sol's output quality?
OpenAI reports a 5.6x end-to-end speedup on the GDP-Val benchmark with no measured quality loss. That's their benchmark, though — confirm it holds on your own prompts before you trust it for anything customer-facing.
Can I use Ultrafast inside ChatGPT, or only through the API?
Only the API for now. OpenAI's announcement covers developer access exclusively; there's no mention of an Ultrafast mode inside the ChatGPT consumer app.