A new Sol, not a promise that every task becomes easy
The September 29 API changelog introduces gpt-6.1-sol for demanding coding and professional work. OpenAI positions it close to Astra at a lower price. That is the vendor's positioning, not an independent finding that the models are interchangeable on every task.
The important upgrade question is local: does the new model finish your work with fewer corrections? For this website, a useful comparison would be preparing a sourced article, updating its metadata, and checking that all its links still work. A better first draft matters, but a publishable result matters more.
Where the new model is available
ChatGPT Learn lists GPT-6.1 Sol in Work and Codex, not standard Chat. The launch includes Plus, Pro, Business, Enterprise, and Edu; Enterprise and Edu need an administrator to enable it. Free and Go are outside the initial rollout. Client versions, rollout status, and workspace controls can still affect the picker.
Standard and Fast were available at launch. OpenAI added GPT-6.1 Sol Ultrafast on October 8 for the Responses API, subject to rate limits. Check the exact model and service tier before interpreting a screenshot or comparing two sessions.
The API settings worth checking
The model reference specifies a 1,050,000-token context window, up to 128,000 output tokens, text input/output, and image input. Audio and video are not supported. Reasoning choices are low, medium, high, xhigh, and max, with medium the default; none and minimal are not supported.
Tool calling requires the Responses API. Chat Completions remains available without tool calling. This distinction matters for an existing integration: changing only the model name does not establish that the surrounding workflow is compatible.
For short-context Standard requests, the changelog lists $2 per million fresh input tokens, $0.10 cached input, $2.50 cache writes, and $10 output. See the dated GPT pricing guide for the other tiers and an example calculation.
A small upgrade test that tells you something useful
- Select ten representative tasks with saved inputs and clear acceptance criteria. Include a routine edit, a reproducible defect, and a task requiring several related files.
- Keep tool access and the supplied material comparable. Record model, reasoning setting, speed mode, and client version so a changed setting is not mistaken for a model improvement.
- Score correctness before style. For code, run the relevant tests; for an article, verify claims, dates, sources, and internal links.
- Count retries, review minutes, and consumed allowance or API cost. Keep failures in the result instead of reporting only the successful runs.
This is our suggested evaluation procedure, not a benchmark we have already run. A model can sound more confident while leaving the same amount of repair work. Conversely, a less polished answer can still contain the correct fix.
Choose the default from accepted results
If GPT-6.1 Sol performs well on the routine set, expand the trial to more ambiguous tasks. Keep an escalation path for work where it repeatedly misses requirements. Preserve the original test cases so later changes can be checked against the same material.
Our GPT-6.1 Sol versus Claude Opus 5.5 comparison puts the API rates, context limits, and evaluation choices side by side. The quality regression checklist covers repeatable comparisons. The subscription usage guide explains why the API price cannot be converted into a fixed number of included tasks.