GPT-5.6 Sol’s Chat update changes the interface, not the Codex model

OpenAI says it updated GPT-5.6 Sol in the standard ChatGPT Chat experience for Plus and Pro users on August 6. The company describes the change as a retuning for everyday conversations: more direct answers, less unnecessary formatting, better adaptation to the requested level of detail, and a willingness to correct a user rather than merely agree. The visible product change is a slider that lets users choose how much reasoning ChatGPT applies to a response. This is an in-product behavior update to an existing model family, not evidence of a new model release.

The timing matters because GPT-5.6 had already reached ChatGPT, Codex, and the API in July as a three-tier family: Sol, Terra, and Luna. In that broader launch, Sol was the flagship tier, while Terra and Luna were positioned as lower-cost and faster alternatives. The August announcement narrows its scope: OpenAI says the refreshed Sol is available only in ChatGPT Chat, and expressly says that the Sol versions serving ChatGPT Work and Codex are not changing. That distinction prevents a product-surface change from being mistaken for a new API or coding-model revision.

For paid ChatGPT users, the current help documentation ties Sol to Medium, High, and Extra High reasoning settings, with Sol Pro reserved for the Pro setting where available. OpenAI also says Free and Go users receive Luna instead of Sol in standard ChatGPT conversations. The practical implication is that the new slider is primarily a control over the Chat experience and its reasoning effort, not a promise that every GPT-5.6-labelled surface will share identical behavior. That is an inference from OpenAI’s stated product split, rather than an independently measured comparison.

Technical users should therefore separate Chat evaluation from API and agent-workflow evaluation. The API documentation still presents GPT-5.6 Sol as the frontier tier, with a 1,050,000-token context window, 128,000 maximum output tokens, and support for tools including web search, hosted shell, computer use, MCP, and skills. Those documented capabilities remain relevant to developers, but OpenAI’s August post does not claim that the Chat retuning changed them. Teams that care about reproducibility should test the exact surface they use and, where appropriate, pin a snapshot instead of assuming a conversational tuning change carries over.

OpenAI reports that, in its internal financial, medical, and legal factual-detail evaluation, responses containing at least one factual error were about 68% less common for the refreshed Sol than for GPT-5.5 Instant. Axios reported the same announcement while identifying the figures as OpenAI’s claims. That makes the result useful context, not independent proof of general accuracy. OpenAI’s help documentation also notes gradual rollout, plan-specific access, usage limits, and safeguards that can affect some legitimate high-risk requests, so availability and observed behavior may vary by account and task.