GPT-5.6 Sol Ultrafast for developers: the eval loop just got 14× shorter (allegedly)

What it is

Ultrafast is a new OpenAI API service tier that runs the GPT-5.6 Sol model on Cerebras hardware instead of standard OpenAI inference infrastructure. OpenAI is previewing Ultrafast, a new service tier that runs GPT‑5.6 Sol up to 14× faster than Standard processing. Powered by Cerebras, Ultrafast can generate up to 750 output tokens per second. Same model weights, different silicon, different queue — that's the whole pitch.

The next-work-session test

You're iterating on a multi-step agent — planner, tool-calling loop, verifier. On Standard Sol, one full trajectory through your eval suite of 200 cases takes ~40 minutes because each case chews through 3–8k output tokens across five turns. You go get coffee. You lose the thread.

At up to 750 output tokens per second, that same suite should complete in the time it takes to read the diff. That's the change: evals become interactive. You edit a prompt, hit run, watch it finish, tweak, repeat — inside one focus block instead of across a day. Same story for latency-sensitive agent loops where each tool call currently stalls the trace.

Whether it delivers 14× on your workload is the question. That number is a peak, not an average.

Pricing

Pricing: unverified. OpenAI has not published Ultrafast rates. Select API customers only, pricing undisclosed, and access is gated — it is currently available as a limited preview to a select group of customers, with access expanding as capacity grows. For reference, Standard Sol runs $5 per million input tokens and $30 per million output tokens. Assume Ultrafast will carry a premium over that; assume nothing about how much. If a rep quotes you a number, that's the number — the public docs don't have one.

What we'd actually use it for

Three narrow things:

What we would not use it for on day one: long-context reasoning traces where the bottleneck is thinking, not emitting, and anything cost-sensitive until pricing is public.

Limits

• Access is gated. Ultrafast mode for GPT-5.6 Sol is now in limited preview. If you're not on the list, this section is theoretical.
• 14× is a ceiling, not a floor. Vendor-claimed peak throughput. Real workloads with tool calls, streaming pauses, and cold starts will land lower. Benchmark before you rewrite anything.
• Single-model tier. Only Sol. No Terra, no Luna, no fine-tunes.
• Vendor lock at the infra layer. You're now dependent on Cerebras capacity availability and OpenAI's queue for it. Two vendors, one path.
• No public SLA, quota, or context-window disclosure for the Ultrafast tier specifically. The OpenAI preview announcement is short on operational detail; the developer forum thread is where questions are being triaged.
• Quality claim is the vendor's. Cerebras powers GPT-5.6 Sol on Ultrafast mode, delivering up to 750 output tokens per second and without any quality compromise — that's Cerebras saying that. Run your own eval before believing it. Weight-identical inference on different hardware has surprised people before (numeric determinism, sampling edge cases).

Try it if

• Your agent loop or eval harness is currently gated by token emission speed, not by reasoning quality.
• You're shipping something latency-facing — voice, live code assist, incident copilots.
• You have OpenAI enterprise contact and can get onto the preview list this quarter.
• You can afford to run a bake-off before committing prod traffic.

Skip it if

• You're on Terra or Luna for cost reasons — Ultrafast is Sol-only and almost certainly more expensive per token.
• Your bottleneck is context length or tool-call round-trips, not raw generation speed.
• You need published pricing to get budget approval.
• Your workload is batch offline — pay Standard rates, wait the extra minutes.
• You're not already an API customer with a path to preview access.