OpenAI ships August refresh of GPT-5.6 Sol and Luna, cuts hallucination rates roughly 60%

OpenAI began rolling out an August refresh of GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT on August 6, 2026, replacing GPT-5.5 Instant as the default models across every ChatGPT tier and adding a new reasoning-effort slider for paid users, according to the accompanying system card.

What's new

"Starting today, we're updating ChatGPT with a more capable model and expanding access for everyone," OpenAI wrote in the GPT-5.6 August Updates system card. "Free and Go users will get a new default model for everyday chats. Plus and Pro users will get an updated GPT-5.6 Sol with a slider that lets them choose how much effort ChatGPT uses for a response. These models will replace GPT-5.5 Instant."

The company draws a line between this release and the versions still running elsewhere: "Users accessing GPT-5.6 Sol and GPT-5.6 Luna in Codex, and via ChatGPT Work, are still using previously released versions of GPT-5.6 Sol and GPT-5.6 Luna. In this system card, we distinguish these models by their month of release: August for the versions released today, and July for the versions that remain in use in Codex and Work."

On factual accuracy, OpenAI reports substantial gains over the outgoing GPT-5.5 Instant. Testing an LLM-based grader against challenging, hallucination-prone prompt sets covering factuality-heavy queries, previously flagged user failures, and high-stakes medical, legal, and financial questions, the company found that "GPT-5.6 Luna reduces factual error rates by over 60% on high-stakes prompts and by roughly 30% on the other two prompt sets," while "GPT-5.6 Sol delivers statistically significant and more consistent improvements overall, reducing factual error rates by roughly 60% across all three prompt sets."

The update also expands OpenAI's youth-safety evaluation regime. "For the first time, we are including dedicated U18 evaluations designed to measure model behavior against teen-specific safety standards," covering "adversarial examples related to self-harm, eating disorder-related behaviors, access to age-restricted goods or services, graphic violent content, and inappropriate sexual content."

Context

GPT-5.6 Sol and Luna have been in production since their initial release, with Sol serving as the higher-capability tier and Luna the lighter, faster option aimed at free-tier usage. This refresh keeps both names but treats the August checkpoints as new models for evaluation purposes, distinct from the July builds still deployed in Codex and ChatGPT Work. It follows a string of recent OpenAI moves to consolidate ChatGPT's default model lineup and retire older Instant-series models, including GPT-5.5 Instant, which this release formally replaces across Free, Go, Plus, and Pro.

Under OpenAI's Preparedness Framework, the company is treating the August release of both models as High capability in the Cybersecurity and Biological and Chemical domains, though neither reaches the High threshold for AI Self-Improvement. OpenAI says it has applied the safeguards outlined in the original GPT-5.6 System Card as a result.

Why it matters

The simultaneous default-model swap for free users and the new effort slider for paid users signal OpenAI's continued push to make ChatGPT's reasoning depth a user-facing dial rather than a background implementation detail, echoing the industry-wide shift toward exposing test-time compute controls. The reported 60% reduction in factual error rates on high-stakes prompts is a meaningful claim given how often hallucinations in medical, legal, and financial contexts have drawn scrutiny of chatbot reliability. Coupled with the new U18 evaluation suite, the release also reflects mounting regulatory and public pressure on AI companies to demonstrate concrete teen-safety measures rather than general policy commitments.