GPT-5.6 Sol’s ChatGPT update is not a Work or Codex model change

OpenAI has updated GPT-5.6 Sol inside ChatGPT for Plus and Pro customers. The company says the change is intended to produce more focused answers and fewer factual mistakes, while letting users choose how much thought a response receives through a new slider. Crucially for developers, OpenAI describes this as a Chat experience update: it says the GPT-5.6 Sol version used by Work and Codex is not changing.

This is a product-behavior revision, not evidence of a new model endpoint or a broader capability change. In its June preview, OpenAI presented Sol as the flagship of the GPT-5.6 family and described high-effort reasoning modes, agentic coding, biology and cybersecurity evaluations, and a phased rollout. The August post narrows its scope to consumer ChatGPT behavior. That split matters because a model name can conceal distinct serving configurations across ChatGPT and developer-facing products.

The measured improvements so far should be treated as provider-reported. OpenAI says internal finance, medical and legal prompts yielded factual-error rates 68% lower for Sol than GPT-5.5 Instant, but it does not publish enough methodology in the announcement to establish an independent benchmark result. An OpenAI-published case study with Model ML gives a different, applied signal: in that company's document-generation harness, Sol completed every PowerPoint task, but trailed some competitors on visual-quality measures and fully correct Excel models. It is useful operational evidence, not a neutral cross-vendor evaluation.

For teams deciding what to change now, the practical answer is conservative. ChatGPT Plus and Pro users can test the slider and compare the new behavior on their own prompts, especially for research, writing, planning and coding; free users get Luna rather than Sol in the described rollout. API, Work and Codex users should not treat this announcement as a reason to assume different production outputs or change routing. The relevant next evidence would be versioned product documentation, API notes, or independently reproducible evaluations, none of which OpenAI identifies in this Chat update.