GPT-5.6 Ultrafast: is a 14× speed preview actually useful in the agent loop?

What it is

Ultrafast is a new OpenAI API service tier — not a new model — that runs GPT-5.6 Sol on Cerebras hardware. 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. Access is gated: Ultrafast mode is available in limited preview to a select group of customers, launching first through the OpenAI API.

Both numbers are OpenAI's claim, measured against its own Standard tier. Nobody outside the preview has benchmarked it yet.

The next-work-session test

You're iterating on an agent that runs a five-step tool loop: plan → search → read → code → verify. Each turn is a Sol call. On Standard, a single eval run over 200 fixtures takes ~40 minutes and you kick it off before lunch. If Ultrafast delivers even a fraction of the claimed speedup on output-heavy turns, that same run finishes while you're still reading the diff.

The concrete change: you stop batching eval runs overnight. You run them per commit. That's the point of caring about this tier at all — not "my chatbot feels snappier," but shorter feedback loops on agent traces you're actually debugging.

Caveat: agent loops are often bottlenecked by tool calls (retrieval, sandboxed code execution, HTTP), not model output. Ultrafast only helps the model-output slice. Measure before you migrate.

Pricing

Unverified for Ultrafast specifically. OpenAI has not disclosed pricing for Ultrafast — the GPT-5.6 Sol pricing page shows the standard rate of $5 per million input tokens and $30 per million output tokens for Standard and Fast mode, but Ultrafast's premium over that, if any, has not been published. For reference, the underlying model is priced at $5 per million input tokens, $30 per million output tokens on the Standard tier via OpenRouter's listing. Treat any Ultrafast quote you see before GA as speculation. See OpenAI's Ultrafast preview announcement for the official post; it doesn't publish a rate card.

What we'd actually use it for

Eval harnesses and CI, in that order.

Point your existing eval runner (Braintrust, Inspect, promptfoo, or your own) at the Ultrafast endpoint for the Sol-model rows. Keep Standard for the "ground truth" baseline row so you can catch quality regressions if Ultrafast's decode path diverges. If you run agent trajectories in CI on every PR, this is where the wall-clock savings show up on the team's dashboard, not in production traffic.

Second use: interactive debugging of a live agent trace. Replaying a stuck loop with fast decode makes it tolerable to step through 20 iterations by hand.

Limits

• Preview access. Select customers only. If you're not already on a named-account track with OpenAI, you're waiting.
• No published SLA or price. You can't put this in a budget spreadsheet yet.
• Speedup is on output tokens. Input processing, tool round-trips, and network hops aren't accelerated. An agent that spends 80% of wall time waiting on a headless browser will see maybe a 20% end-to-end win.
• Provider-specific decode path. Cerebras inference can produce subtly different token distributions from GPU inference on the same weights. Re-run your regression evals — don't assume parity.
• Terminal-Bench numbers are Sol, not Ultrafast. On Terminal-Bench 2.1 (agentic coding), Sol Ultra scored 91.9% and base Sol scored 88.8%. Ultrafast is a tier for the same model; verify quality on your own evals before assuming those scores carry over.
• Target workloads are broader than agents. The new API service tier is powered by Cerebras and targets coding, financial research, customer support, commerce — meaning the vendor optimizes for a mix, not specifically your agent loop.

Try it if

• You already have a working eval harness against GPT-5.6 Sol and iteration speed is your bottleneck.
• Your agent's wall-clock is dominated by model output, not tool calls.
• You have direct account contact with OpenAI to request preview access.
• You can A/B test Ultrafast vs Standard on quality before cutting production over.

Skip it if

• You're on Terra or Luna. This preview is Sol-only.
• Your agent spends most of its time in retrieval, sandboxes, or third-party APIs — speeding up decode won't move your p95.
• You need firm pricing to justify the switch to finance.
• You need it in the next sprint. Preview access timelines aren't published, and expanded access to more businesses as capacity grows is not a date.