OpenAI cuts ChatGPT errors 68% and gives free users unlimited Luna

OpenAI is giving away more than it used to charge for. On 7 August 2026 it rolled out GPT-5.6 Luna as the default model for free and Go users, with unlimited text chats, alongside an updated GPT-5.6 Sol for paid tiers that OpenAI says makes far fewer factual errors. The move tightens OpenAI's grip on consumer usage at the same time as it narrows the gap between free and paid quality.

The headline figure is accuracy. On OpenAI's own tests across financial, medical and legal prompts, responses containing at least one factual error were 68% less common with the updated GPT-5.6 Sol and about 62% less common with GPT-5.6 Luna, compared with GPT-5.5 Instant. OpenAI attributes the gains to better use of the sources the model looks up. These are internal figures, not an independent benchmark.

For Plus and Pro subscribers, OpenAI has merged its fast and reasoning modes into a single Sol model, with a slider to control how much reasoning it applies. Free and Go users get a simpler "Think" button instead. The model picker has moved into settings, a change that drew complaints on X.

The update is scoped to ChatGPT's consumer chat product; Sol in ChatGPT Work and Codex is untouched, and OpenAI is calling this a refresh of GPT-5.6 rather than a new model generation. OpenAI also said it strengthened protections for under-18s, including tighter limits on romantic role-play, according to DataNorth.

The timing matters. Anthropic has the lead in business spending on AI, so OpenAI's counter-move is to entrench itself where it is already strongest: consumer reach. Unlimited free chats on Luna raise the cost of switching for casual users and widen the funnel toward paid tiers, while the accuracy claims target the credibility gap that has kept some professional users on rivals or on manual fact-checking.

The risk is margin. Giving away unlimited access to a model OpenAI is touting for factual reliability implies real inference cost with no direct revenue from that tier. Watch whether OpenAI's paid conversion improves enough to offset it, and whether independent testing backs the 68% error-reduction claim once outside researchers get access.