OpenAI published a GPT-5.6 August update on 2026-08-06 covering revised versions of GPT-5.6 Sol and GPT-5.6 Luna and a change to what ChatGPT serves by default. Free and Go users move to a new default model for everyday chats, while Plus and Pro users get an updated GPT-5.6 Sol with a slider that controls how much effort ChatGPT spends on a response. Both replace GPT-5.5 Instant.
The reported gains are concentrated in factuality. Against GPT-5.5 Instant, OpenAI reports roughly a 60 percent reduction in error rate on its Factuality-Heavy set, roughly 60 percent on its High Stakes set, and roughly 30 percent on a set built from user-flagged failures. On health evaluations, GPT-5.6 Sol scores 55.0 on HealthBench, 31.4 on HealthBench Hard and 54.0 on HealthBench Professional, all length-adjusted.
The document also records the Preparedness Framework classification for both models: High capability in the Biological and Chemical domain, High capability in Cybersecurity, and below High in AI Self-Improvement. OpenAI notes a regression on a self-harm metric in its mental health evaluations, while stating that online testing did not show an increase in problematic responses.
The operational detail worth acting on is the default swap. Any organization whose employees use free or Go ChatGPT has had the underlying model changed without a migration step, and any evaluation harness pinned to GPT-5.5 Instant is now measuring a model that consumer users no longer see. The dual High designation in biological, chemical and cybersecurity domains is also a governance data point: OpenAI is now shipping a mass-market default that it classifies as high capability in two of its most safeguarded categories.