GPT-6.1 Sol Ultrafast costs $12/$60: pay only where users wait

OpenAI is rolling out GPT-6.1 Sol Ultrafast in the API, Codex and ChatGPT Work, and the price is now on the page: $12 per million input tokens and $60 per million output. That is six times what the same model costs on the Standard tier. OpenAI says you get "near-Astra intelligence at up to 8x faster speeds." Speed is now a line item. The question for your agents is not "is it faster?" It is "which step is worth 6x?"

What actually happened with GPT-6.1 Sol Ultrafast

OpenAI announced the rollout on its developer forum this week. The details from the Ultrafast docs and the pricing page:

• Price per million tokens (up to 272K input): gpt-6.1-sol costs $2 / $10 on Standard, $4 / $20 on Fast, and $12 / $60 on Ultrafast. Cached input is $0.60 on Ultrafast.
• Astra is worse. GPT-6 Astra Ultrafast lists at $60 input and $300 output. OpenAI's own pitch is that Sol Ultrafast costs "just 1.2x" what Astra costs on Standard.
• How to turn it on: set service_tier to ultrafast on each request. OpenAI strongly recommends WebSockets for agent loops with many short tool calls, because HTTP overhead eats the gain.
• Separate rate limits: Sol Ultrafast starts at 1M tokens per minute on the Build tier, 4M on Launch and 40M on Grow. These limits are separate from Standard and Fast.
• Data residency: Sol Ultrafast supports US and EU residency. Astra Ultrafast is US only.
• Seats: in Codex and ChatGPT Work, Ultrafast is on Pro 500, eligible usage-based Enterprise and credit-based Edu plans. Enterprise admins must enable it.

The "up to 8x" speed figure is OpenAI's claim. The docs give no tokens-per-second number for the API.

Why it matters for your business

We said this when Ultrafast first appeared on GPT-5.6 Sol: faster tokens do not make your database, your payment API or your retries faster. If model generation is 30% of your agent's wall clock, an 8x model speedup saves about a quarter of the total time. You pay 6x on every token for it.

So split the work by who is waiting:

• A person is waiting: a live chat answer, a support agent on a call, an operator stuck on an outage. Ultrafast can pay here.
• Nobody is waiting: overnight catalog enrichment, invoice extraction, report drafts. Use Standard, or the Batch API. Speed you cannot see is money you lose.
• Cache first. A cached input token on Ultrafast costs $0.60, against $12 uncached. A stable prompt prefix cuts the premium more than any other change.

Sources: OpenAI Developer Community, OpenAI Ultrafast docs, OpenAI API pricing.