GPT-6.1 Sol reaches the API, Work and Codex

GPT-6.1 Sol gives developers another candidate for coding and professional work when Astra’s cost is difficult to justify. OpenAI’s September 29, 2026 changelog confirms the API release as gpt-6.1-sol. Its description of near-Astra performance is a vendor claim, not a universal result from our testing. We checked the current documentation on October 4. OpenAI API changelog, model documentation.

API access and subscription rollout are separate

The same date’s ChatGPT notes describe a rollout in Work and Codex, starting with Pro and expanding to Plus, Business, Enterprise and Edu. Managed workspace permissions can affect access. That does not mean every account receives the model simultaneously, or that a subscription pays for developer API requests. ChatGPT release notes.

For API integrations, use Responses for tool calling. Chat Completions is supported without tools. The documented reasoning levels run from low through max, with medium as default; none and minimal are unsupported. Check these settings before changing an existing request. Sol model requirements.

The practical distinction is the software around the model. Choosing Sol in a coding app changes a configured assistant. Calling Sol from your own application also leaves you responsible for its tools, execution environment and acceptance checks. The models, apps and agents explainer separates those choices.

Price a complete attempt

For Standard API requests with at most 272K input tokens, the checked rates per million tokens are $2 input, $0.10 cached input, $2.50 cache writes and $10 output. Above that threshold, the full request uses double input/cache rates and 1.5 times the output rate. Processing tiers and regional premiums can change the bill. OpenAI pricing, model pricing details.

Here is an illustrative calculation, excluding caching, tools and other charges: 100,000 input tokens plus 20,000 output tokens at those Standard rates cost $0.20 + $0.20 = $0.40. Three equal attempts would cost $1.20. These are arithmetic assumptions, not observed usage.

For an agent, add the usage from every call, including unsuccessful work. Then divide the total by accepted deliverables. A low token bill is not enough if reviewing and repairing the result takes longer than your existing workflow.

Compare one job before switching

Our recommendation is a bounded comparison on work you can grade. For example, ask Sol and your current assistant to fix the same reproducible form bug, using the same repository and instructions. Require the form submission, validation and mobile behavior to work. Record failures, total spend and your review time; do not judge only the persuasiveness of the completion message.

If Sol meets the checks at a lower overall cost, expand the trial to a different task. If it repeatedly fails a difficult case, retain the stronger model for that case. No independent results in this report establish one default for everyone.

Next action: confirm account access or API settings, then evaluate one inspectable task. Use the Codex workflow guide for practical acceptance checks, and the Astra release explainer for the higher-cost alternative’s scope.