OpenAI's Ultrafast GPT-6.1 Sol runs up to 8x faster at a premium. Pay for it only where someone is waiting
Ultrafast for GPT-6.1 Sol is rolling out in the API, Codex and ChatGPT Work at $12 and $60 per million tokens. The case for it is wall-clock time in loops and live work, not speed for its own sake.
Listen to this article · 4 min
AI-generated narration of the full article.
When is an answer worth paying more for because it arrives sooner?
OpenAI put a price on that question on October 8. Ultrafast mode for GPT-6.1 Sol is rolling out in the API, Codex and ChatGPT Work, offering what OpenAI calls “Near-Astra intelligence at up to 8x faster speeds” than Sol in its standard mode. In the API it costs $12 per million input tokens and $60 per million output tokens, which OpenAI describes as “just 1.2x the cost of Astra”, its top reasoning model. When we covered OpenAI’s GPT-6 model guide earlier this month, Ultrafast was available only for Astra.
The details that matter for enterprise planning: in Codex and ChatGPT Work, access is limited to Pro 500, eligible usage-based Enterprise and credit-based Edu plans, and Enterprise admins must enable it. In the API, Ultrafast Sol is available in all supported regions, including data residency in the U.S. and the EU, and OpenAI’s developer guide describes using it over WebSockets. OpenAI positions it for debugging outages, agents navigating apps and live experiences.
Speed compounds in loops
A faster model matters little for a single question. It matters a great deal for an agent that makes forty model calls to finish one task, because every saved second is saved forty times, and the person waiting experiences the total. That is why OpenAI’s chosen examples are an engineer debugging an outage and an agent navigating applications: in both, wall-clock time is the cost that hurts, and the token bill is secondary.
The same logic cuts the other way. A nightly reconciliation finishes before anyone looks at it, whatever the speed. Paying a premium there buys nothing.
| Workload | Who waits | Ultrafast? |
|---|---|---|
| Debugging a live incident | An engineer, under pressure | Strong case |
| Agent navigating applications | A loop of many steps | Test it: savings compound per step |
| Voice or live customer interaction | A customer, in real time | Strong case, if quality holds |
| Background agent reporting later | Nobody, until the result | Usually not |
| Overnight batch or reconciliation | Nobody | No |
Three checks before turning it on
Measure your own speed and quality. “Up to 8x” is a ceiling, and “near-Astra intelligence” is OpenAI’s framing. Run your evaluation set and record p50 and p95 latency on your real prompts. The unit to compare is cost per successful task and time to result, not price per token.
Make it a policy, not a toggle. Enterprise admins must enable Ultrafast in Codex and ChatGPT Work. Use that gate to decide which teams and workflows get it and why, rather than turning it on for everyone. In the API, set the mode per workflow step, with the same discipline we described for the price line in Claude Haiku 5.5: routing decides most of the bill.
Put a budget next to it. A premium mode used in an agent loop multiplies quickly. Spend alerts by workflow, the kind of control we covered in AI spend guardrails, should be in place before the first incident team discovers how much faster its agent feels.
Ultrafast is a useful option, and in incident response it may pay for itself in the first hour of an outage. It is also the easiest premium in the stack to overuse, because faster always feels better. Pay for speed where a person or a loop is waiting, and nowhere else.