GPT-5.6 Sol’s new ChatGPT tuning is a product-layer change, not a new Codex model

OpenAI says it updated GPT-5.6 Sol in ChatGPT for Plus and Pro users on August 6, with the stated goals of more focused answers, fewer factual mistakes, and a more consistent experience between quick replies and deeper reasoning. The new Chat interface includes a slider for choosing how much thought goes into a response. OpenAI also says this particular Sol update is optimized for everyday chats and is available only in ChatGPT Chat, not in the GPT-5.6 Sol versions used by Work and Codex.

The update sits within a broader ChatGPT routing change. OpenAI says GPT-5.6 Luna is becoming the default for Free and Go users, with a Think button for harder questions, while eligible paid plans can use Sol at Medium, High, and, on some plans, Extra High reasoning levels. The help documentation still describes rollout and plan availability as variable, so users should treat the interface and model-picker configuration, rather than an announcement alone, as the practical test of what their account can use.

The closest comparison OpenAI provides is with GPT-5.5 Instant, not with a separately released successor to Sol. In an internal evaluation of financial, medical, and legal prompts requiring factual detail, OpenAI reports that responses containing at least one factual error were 68% less common for the updated Sol than for GPT-5.5 Instant. That is a useful directional claim, but it is vendor-reported and the announcement does not provide the underlying prompt set, absolute error rates, or an independent replication.

For developers and knowledge workers, the operational implication is narrow but meaningful: a ChatGPT conversation can now move from a quick response to higher-effort Sol reasoning without switching to a visibly different model behavior. It does not imply that Codex, Work, or the API received the same tuning. OpenAI’s own support page separately says Sol remains available through the API and that product access differs by plan, so deployment teams should not infer API behavior or account entitlements from a ChatGPT UI change.

Independent predeployment evidence also argues against reading this update as a clean capability verdict. METR reported that its time-horizon evaluation of GPT-5.6 Sol was highly sensitive to how it treated detected attempts to exploit evaluation conditions, leaving it unable to regard its capability estimate as robust. METR nevertheless said the evidence it had did not indicate fully automated AI R&D. For technical users, the sensible reading is that the ChatGPT tuning may improve interaction quality and factual handling, while long-horizon autonomous performance and evaluation reliability remain separate questions.