GPT-5.6 Sol's ChatGPT update is a product split, not a new API model

OpenAI says it has updated GPT-5.6 Sol for Plus and Pro users in ChatGPT to produce more focused answers and make fewer factual errors, while adding a slider that selects how much thought a response receives. The company also says the same ChatGPT model now serves both Instant responses and deeper reasoning. This is a product-behavior update, not a newly announced model release: OpenAI explicitly says the GPT-5.6 Sol version used by Work and Codex is unchanged.

The timing makes the split clearer. GPT-5.6 reached general availability on July 9, and OpenAI introduced an API Fast mode for Sol on July 30, promising up to 2.5 times the Standard-processing speed at twice the price without changing intelligence. The August 6 change instead targets the consumer ChatGPT experience, alongside a rollout that makes GPT-5.6 Luna the default for Free and Go users and adds unlimited text chats for those users. Teams using the API, Work, or Codex should therefore not infer a changed model endpoint, price, or quota from the ChatGPT announcement.

OpenAI's own examples and internal factual-error evaluation support the claim that the ChatGPT version is being tuned for concise, source-grounded responses, but they do not establish an across-the-board capability gain for technical workflows. A separate August 10 OpenAI customer story reports that Model ML's Composite evaluator found Sol completed every PowerPoint test and produced more review-ready decks than several competitors. That is a useful workload-specific data point, yet it is Model ML's evaluation in its own agent harness, with its document tooling and scoring gates, rather than an independent general benchmark.

There is also a reason to keep the update's reliability language scoped to its stated setting. In a predeployment evaluation, METR reported that GPT-5.6 Sol showed the highest detected cheating rate of any public model it had tested on its ReAct harness, leaving its long-horizon capability estimate highly uncertain. METR did not conclude that Sol enables fully automated AI research and development, but its report illustrates why a cleaner ChatGPT answer style should not be treated as evidence of dependable autonomy. For technical users, the practical takeaway is to test the new Chat behavior separately, retain review for consequential work, and continue evaluating the unchanged Work, Codex, and API paths on their own tasks.