At DevDay 2026 on 29 September, OpenAI released GPT-6.1 Sol — a model priced at exactly one-fifth of GPT-6 Astra's rates, with results close to the flagship. A new Ultrafast tier launched alongside it, pushing throughput to 300 tokens per second. The product here is a lower bill, not a higher ceiling.
Key takeaways
- $2 per million input tokens and $10 per million output tokens
- Astra costs $10 and $50 — Sol is exactly one-fifth of that
- Cached input dropped from $0.20 to $0.10 per million tokens
- Ultrafast: up to 300 tokens per second, 6x faster in the API, 8x in Codex
- The Ultrafast tier costs six times the standard rate
The same work, one-fifth of the bill
The ratio is exact rather than rounded — it holds for input and output tokens alike.
Symbol meaning
- …
- GPT-6.1 Sol, one million input tokens
- …
- GPT-6 Astra on input, and Sol on output
- …
- GPT-6 Astra, one million output tokens
| Line item (1M tokens) | GPT-6.1 Sol | GPT-6 Astra |
|---|---|---|
| Input | $2 | $10 |
| Output | $10 | $50 |
The strongest result is in code: on DeepSWE v1.1 Sol matches Astra and beats the older GPT-6 Sol by 6.4 percentage points. On enterprise tasks in AutomationBench it leads Claude Opus 5.5 by 2.2 percentage points and its predecessor by 4.8.
In computer use (OSWorld 2.0) the gap to Astra is 2.1 percentage points at maximum reasoning effort, at one-seventh of the cost. Cached input halved in price, which matters for agents replaying the same context.
Ultrafast: speed as a separate line item
Ultrafast is not a new model but a faster way to run one. OpenAI cites up to 300 tokens per second — 6x faster in the API and 8x in Codex. The price rises sixfold. The tier currently runs on Astra, for Pro 500 and Enterprise customers.
In that line-up Ultrafast is not the fastest option on the market — it is the fastest in OpenAI’s own range. Mercury 2 stays ahead, Gemini 3.5 Flash falls behind.
| Model / tier | Tokens per second |
|---|---|
| Mercury 2 | 769 |
| GPT-6 Astra on the Ultrafast tier | up to 300 |
| Gemini 3.5 Flash | 201 |
Where Sol lands, and where it does not
The model arrives in the API and in ChatGPT Work, Codex and the Plus, Pro, Business, Enterprise and Edu plans. Regular Chat does not get it. The direction is clear: Sol is meant to work, not to converse.
Why it matters
Price has stopped tracking capability. If the gap to the flagship fits inside two percentage points, the case for the pricier variant shrinks to edge cases. Splitting speed and intelligence into separate line items moves the decision to the system architect: what needs 300 tokens per second, and what can wait. That changes agent budgeting more than another index point would.
What next?
- Ultrafast for GPT-6.1 Sol is announced, with no date given
- Astra Ultrafast is available today only to Pro 500 and Enterprise customers
- Sol stays out of regular ChatGPT Chat, with no change announced
Sources
- VentureBeat — OpenAI's GPT-6.1 Sol offers Astra-like performance at 1/5th price
- Simon Willison's Weblog — OpenAI DevDay 2026 live blog
