OpenAI's GPT-6 October Update Becomes the ChatGPT Default, With Documented Safety Regressions
OpenAI has replaced GPT-5.6 Sol and Luna with GPT-6 Sol and GPT-6 Luna as the default models in ChatGPT, across free and paid plans, and published the safety tradeoffs alongside the release. The October update documentation is the useful read for anyone whose team works inside ChatGPT daily.
The headline numbers are mostly improvements. Instruction hierarchy adherence is reported at 99.99 percent for Sol and 99.79 percent for Luna. On HealthBench Professional, Sol scored 55.2 versus 54.0 for GPT-5.6 Sol, and Luna scored 48.2 versus 44.1. Auto-review recorded prohibited actions in 0 percent of rollouts, down from 0.3 percent for the August models.
The regressions are the part to read closely. Multiturn jailbreak resistance is slightly lower than the September models, though both new models still beat GPT-5.6 Sol. Sol regressed on standard self-harm content; Luna regressed on self-harm, gore and sexual content; both regressed on under-18 content, which OpenAI mitigates with a classifier-based block. The company treats both models as High capability in Cybersecurity and in Biological and Chemical domains, but not High in AI Self-Improvement. ExploitBench scores were 82.62 percent for Sol and 44.66 percent for Luna, both below the Critical threshold.
One operational detail matters for rollout: Codex and ChatGPT Work still use the September versions. A company can therefore have people on different model generations at the same time, depending on which surface they open.
For leaders running AI adoption, the takeaway is that a default-model change arrives without a project plan. Know which model each workflow runs on, re-test the prompts and approvals that matter after every swap, and keep the human checks in place where the regression list touches your use case. Outcomes depend on the version actually running, not the one you evaluated last quarter.