GPT-5.6 Sol’s new ChatGPT behavior is intentionally separate from Codex

OpenAI’s August 6 update changes GPT-5.6 Sol inside ChatGPT for Plus and Pro users. The company says the Chat version is being tuned for more focused answers, stronger factual reliability, and more consistent behavior between quick responses and deeper reasoning. It also adds a slider on web, mobile, and desktop that lets those users choose how much thought ChatGPT applies to a response. This is a product-behavior update to an existing model family, not evidence of a new detector-confirmed model release.

The company frames the change as a response to a broad conversational workload: short questions, planning, research, writing, and decisions. Its examples emphasize answering the user’s immediate question first, reducing unnecessary formatting, and carrying context into a follow-up rather than repeating the full answer. OpenAI also reports an internal evaluation in which factual-error responses on financial, medical, and legal prompts were less common for GPT-5.6 Sol than for GPT-5.5 Instant. That is vendor-reported evaluation evidence, so it describes OpenAI’s measurement rather than an independently reproduced result.

The most consequential detail for technical users is the scope boundary. OpenAI explicitly says this everyday-chat-optimized version of GPT-5.6 Sol is available only in ChatGPT, while the GPT-5.6 Sol version powering Work and Codex is not changing in this update. That distinction follows a July product split in which OpenAI described lower prices for Luna and Terra, while presenting faster Sol API processing through Fast mode. A team that uses both ChatGPT and Codex should therefore avoid assuming that observed changes in chat tone, brevity, or effort controls transfer to its coding or API workflows.

The update also reshapes the lower end of the consumer product line. OpenAI says GPT-5.6 Luna will become the default model for Free and Go users, with unlimited text chats and a Think button for harder questions, while limits remain for file uploads, images, and other tools. In other words, the expansion is not a blanket removal of product limits. It is a specific change to text-chat access and an optional higher-reasoning path, subject to the company’s abuse guardrails.

For practitioners, the practical response is to treat the ChatGPT update as a separate surface to evaluate. The new effort slider may make it easier to select faster or more deliberative answers within ChatGPT, but it does not replace task-level evaluation of API latency, tool use, output structure, or Codex behavior. OpenAI’s claims about improved factual reliability and consistency are useful hypotheses for those tests, not a substitute for them, and the announcement leaves the underlying training changes and external benchmark replication unspecified.