Correction, October 13, 2026: Cached-input prices apply when a request reuses an eligible cache entry. Repeating text does not guarantee a cache hit.
What changed
OpenAI released GPT-6.1 Sol on September 29 as an upgrade to GPT-6 Sol. OpenAI says it approaches GPT-6 Astra on multi-step coding tasks and computer use, where an agent operates software through its interface. Its standard API input and output prices are one-fifth of Astra’s.
OpenAI’s API lets an app send requests directly to its models. It bills text in tokens, small chunks of text. GPT-6.1 Sol’s standard API rates are $2 per million input tokens and $10 per million output tokens, the same as GPT-6 Sol. When a request reuses an eligible matching cache entry, OpenAI bills those tokens at the cached-input rate. That rate is now $0.10 per million tokens, half the earlier Sol rate of $0.20 listed in the API changelog.
GPT-6 Astra costs $10 input and $50 output per million tokens, according to OpenAI’s model catalog. GPT-6 Astra targets long coding tasks, while Sol and Luna offer lower-cost tiers.
OpenAI reports that GPT-6.1 Sol matched Astra on DeepSWE v1.1, a test of long software-engineering tasks, at about one-fifth of the per-task cost. Results depend on the task and agent setup.
OpenAI’s reported DeepSWE v1.1 results from its launch announcement. Each point pairs a score with estimated API cost per task at a different reasoning setting. Open the chart to read it at full size.
Why it matters
For builders sending requests from their own app, the meaningful change is more capability at the same standard token rate as GPT-6 Sol. The model ID is gpt-6.1-sol. OpenAI also lists it in Codex and ChatGPT Work.
Check OpenAI’s model migration guidance before swapping models. GPT-6.1 Sol accepts reasoning-effort settings from low to max, but not none or minimal, two lower-effort settings supported by GPT-6 Sol. For tool calls, OpenAI directs developers to the Responses API, its request format for letting a model use tools.
What builders should do next
Pick a coding task with a clear acceptance check, such as fixing a failing test in a non-critical part of your app. Run it on GPT-6 Sol and GPT-6.1 Sol from the same code snapshot, with the same prompt and reasoning setting. Check whether the required test passes, how many correction prompts each run needed, and the API tokens each run used.
If GPT-6.1 Sol passes the same acceptance check with fewer corrections or lower measured usage, compare it on other common tasks before changing your default. Keep Astra for work where your own comparisons show its extra capability is worth the higher token price. If you choose Astra for coding-agent work, reconsider the AGENTS.md and skills instructions written for older models.
End of article