OpenAI is splitting ChatGPT’s model lineup along subscription lines this week. Plus and Pro subscribers get an upgraded GPT-5.6 Sol and a new effort slider starting today, while Free and Go accounts are set to have their default model swapped to GPT-5.6 Luna over the coming days. The two updates ship on different clocks, and only one of them is live right now.
The stakes are commercial as much as technical. ChatGPT serves roughly 1 billion people weekly, according to OpenAI’s own figures, which makes a default-model change one of the largest single deployment decisions in consumer software. Whichever model sits behind that default button shapes the baseline experience for the overwhelming majority of ChatGPT’s user base, long before anyone opens a settings menu to compare options.
OpenAI describes this as broadening access to its newest models. That framing holds: Luna is a more recent model than what free accounts previously defaulted to, and the company says unlimited text chats will follow for Free and Go users starting next week, alongside a Think button they can tap when a question needs more reasoning, pending unspecified abuse guardrails. Neither of those two features is available yet. A second, equally accurate reading is that OpenAI is routing its largest, least monetized user segment onto the cheaper model in its lineup while keeping Sol, the more capable one, gated behind a subscription. Both things are true at once: access is expanding, and it is expanding toward the weaker tier.
GPT-5.6 Sol, the model available to Plus and Pro users today, is scoped narrowly. It only powers the Chat surface; the version of Sol running inside OpenAI’s Work and Codex products is unchanged by this release. OpenAI says the update reduces unnecessary formatting, adjusts response length to the question, and merges what were previously distinct Instant and Thinking behaviors into one model reachable through the new slider.
On accuracy, OpenAI reports that in an internal evaluation covering financial, medical, and legal prompts, the share of answers carrying one or more errors of fact fell by roughly 62 percent with Luna and 68 percent less often with Sol, compared with the outgoing GPT-5.5 Instant. Both figures come from OpenAI’s own testing, run on prompts OpenAI selected, scored by OpenAI’s own evaluation process. That is meaningfully different from a third-party benchmark or an independent audit, and it says nothing about how either model performs on the far larger and messier set of questions people actually ask outside those three domains. A vendor-graded drop in error rate on curated prompts is evidence of improvement, not a reliability guarantee for open-ended use.
The rollout also carries new safety scaffolding aimed at users OpenAI believes are minors, including restrictions on romantic roleplay and boundaries around sensitive content categories, detailed in the release’s accompanying system card.
For builders evaluating ChatGPT as a benchmark or fallback model, the practical takeaway is to test against Sol, not Luna, since Sol is what OpenAI is positioning as its more capable consumer-facing model this cycle. For anyone tracking OpenAI’s free-to-paid conversion funnel, watch whether unlimited chats and the Think button actually land next week as promised. A slipped date on the free tier’s headline feature, arriving days after the default model itself already switched, would be worth noting on its own.
Announced by OpenAI on August 6, 2026.
