Michael JohnsonOn September 29, 2026, OpenAI released GPT‑6.1 Sol for the API. With standard processing, one million input or output tokens for this model costs five times less than for GPT‑6 Astra. But a lower rate does not prove equivalent performance, nor does it mean every developer task will cost 80% less.
OpenAI describes Sol as a model with performance close to Astra on complex tasks, including coding and computer use. For now, this is the company’s positioning, not proof of parity on everyday development tasks.
How much cheaper? Comparing API rates
| Metric, per 1 million tokens | GPT‑6.1 Sol | GPT‑6 Astra | Difference |
|---|---|---|---|
| Input, standard mode | $2 | $10 | Sol is 80% cheaper |
| Cached input | $0.10 | $1 | Sol is 90% cheaper |
| Output, standard mode | $10 | $50 | Sol is 80% cheaper |
| Context | 1.05 million tokens | 1.05 million tokens | Limits match |
| Maximum output | 128,000 tokens | 128,000 tokens | Limits match |
These are API rates for standard mode on requests with up to 272,000 input tokens. Higher rates apply to the entire request for longer inputs; Fast, Batch, and Flex modes also change the price. Tool calls may be billed separately. So the table compares token costs, not the full cost of completing a task. These terms are listed in the GPT‑6.1 Sol model card.
What changes for developers
The API model identifier is gpt-6.1-sol. For tool calls, OpenAI recommends the Responses API; Chat Completions is supported without tools. The model accepts text and images, but not audio or video. Fine-tuning is not supported.
When migrating an existing integration, pay attention to the reasoning.effort settings: the supported values are low, medium, high, xhigh, and max, but not none or minimal. If your client uses unsupported values or calls tools through Chat Completions, you’ll need to change the configuration.
A practical next step is to compare Sol and Astra on the same tasks from your own repository: fixing bugs, writing tests, and making changes across multiple files. Evaluate patch quality after review, the number of retries, token use, time, and tool errors. This is a proposed evaluation method, not a result from a published test.
Sol and Astra: what we know—and what we don’t
| Criterion | What can be concluded |
|---|---|
| Positioning | OpenAI says Sol is close to Astra in capability on complex tasks, but costs less. |
| Rates | Sol’s standard input and output token rates are 80% lower; cached input is 90% cheaper. |
| Context and maximum output | The published limits for both models match: 1.05 million and 128,000 tokens, respectively. |
| Performance on development tasks | Universal parity has not been established; results need to be tested on your own tasks. |
Published Astra evaluations cannot be treated as Sol results. The materials reviewed contain no comparable independent test showing how the two models handle the same development tasks with identical settings and costs.
A brief timeline
- September 22, 2026 — OpenAI announced GPT‑6 Sol and GPT‑6 Luna; this was a separate event, not the launch of GPT‑6.1 Sol, as shown in the announcement to developers.
- September 29, 2026 — The API changelog added an entry announcing the release of GPT‑6.1 Sol. The Associated Press also covered the DevDay announcement.
- September 30, 2026 — OpenAI’s documentation contains the model’s specifications and rates; early user reviews remain too scattered to serve as a comparative test.
Early user impressions do not provide a controlled comparison: authors use different tasks, settings, and methods for calculating costs.
In short, Sol’s confirmed advantage is lower API rates. Developers will have to find out on their own code how much this saves on a particular task—and whether quality holds up.