GPT-6.1 Sol arrived on September 29, one week after GPT-6 Sol and Luna. OpenAI positions it between the cost-sensitive Luna route and its most capable Astra model. For a coding team, the interesting question is whether it can finish your normal multi-file work at an acceptable cost—not whether a new version number deserves an automatic upgrade. Release source.
Verified September 29, 2026. This is a documentation-based upgrade guide. Agent.Space has not run a head-to-head benchmark for this article. Provider API prices below are separate from subscription allowances and Agent.Space prices.
GPT-6.1 Sol API pricing
The Standard API rates below are in US dollars per million tokens, for requests with up to 272K input tokens. OpenAI's pricing page separates Standard, Batch, Flex, and Fast processing; a discounted table is not the Standard rate.
Keep these usage categories separate when estimating an agent run. Repeated repository context can include cache reads, new uncached input, and cache writes; generated output is another meter. Multiplying the whole transcript by the cheapest column produces a misleading budget.
For example, consider an illustrative request, not a measured agent task: 100,000 uncached input tokens plus 10,000 billable output tokens, with no cache writes, tools, or other charges. At those rates, Sol costs $0.30, Luna $0.015, and Astra $1.50. The arithmetic describes the same token volume; it does not establish that the models need the same number of calls to produce an accepted patch.
The model specification lists a 1.05M context window and 128K maximum output. Above 272K input tokens, the full request uses doubled input/cache rates and 1.5× output rates. A large context window is therefore both a capacity option and a billing decision. It does not establish the usable context of every Codex client.
Upgrading from GPT-6 Sol changes more than the model ID
The most consequential compatibility differences are in OpenAI's GPT-6 migration guidance:
- GPT-6.1 Sol supports
low,medium,high,xhigh, andmax; it does not supportnoneorminimalreasoning. - Tool calling requires the Responses API. Chat Completions supports requests without tools.
- Requests using reasoning need an audit of sampling and log-probability parameters; do not carry incompatible settings across unchanged.
An existing GPT-6 Sol integration using Chat Completions tool calls with reasoning_effort: "none" therefore needs a protocol and parameter review. Replacing the ID alone can break the workflow even if a plain-text smoke request succeeds.
Before changing a saved default, exercise one tool call, one multi-step task, and your normal failure handling. Check that the application receives the tool result, continues correctly, and records usage under the intended account. Do this in a controlled environment before moving production traffic.
Choose GPT-6.1 Sol in Codex
On a supported Codex CLI and an account with access, start a task with:
The Codex model page documents this ID and the separate rollout. Access still depends on the client, plan, and workspace settings. If the option is missing, follow the Work and Codex model-picker checklist before changing billing or configuration.
Decide with three tasks from your own backlog
Our suggested trial has three parts: a reproduced bug, a feature crossing an existing interface, and an investigation with a known answer. Start each candidate from an independent copy of the same repository revision. Keep the brief, permissions, tools, and acceptance checks comparable.
Record accepted results, full usage, elapsed time, and reviewer corrections. A lower token bill matters only if the result is usable. If a cheap route needs repeated repairs, that belongs in its total cost. If a stronger route prevents an expensive mistake, that belongs in the decision too. The coding-agent model selection guide explains how to turn this evidence into a working default.
Bring the task to Agent.Space
Agent.Space lets you choose a supported Agent harness and a compatible model around a saved cloud project. First check current model pricing and the live selector: this release does not establish that GPT-6.1 Sol is available for your account or harness in Agent.Space.
When the combination you need is available, start an Agent.Space Workspace with one bounded task and keep the files, acceptance checks, and review notes with the project. For a server integration, use the Developer API setup guide and discover the exact supported model ID rather than copying OpenAI's ID into a different provider route.