OpenAI sharpens GPT-5.6 Sol in ChatGPT, unlimited free chats on Luna
The Chat-only tune-up cuts factual errors 68% but leaves the API untouched; free tier gets a Think button, as Qwen3.8 Max tops the agentic index.
Copy markdown
Sol got sharper — but only inside ChatGPT
OpenAI says GPT-5.6 Sol now returns 68% fewer factual errors in finance, medical, and legal, with tighter, more direct answers. Crucially, the Sol powering the API, Codex, and Work is unchanged — your integrations behave exactly as before.
Free and Go tiers now default to Luna, unlimited
GPT-5.6 Luna becomes the default for free and Go users this week, with previously rate-limited text chats now unlimited — a zero-cost way to prototype against OpenAI's small model. File uploads, image generation, and other tools stay capped.
A Think button, not a bigger brain
Free users get a Think button for hard questions, but it only lets Luna reason longer — it never escalates to a stronger model. Plus and Pro instead gain a five-setting reasoning slider, already shipping in ChatGPT Work.
Elsewhere: Qwen3.8 Max tops the agentic index
Alibaba's Qwen3.8 Max now leads Artificial Analysis's Agentic Index at 55.4, ahead of every US lab except the top Anthropic and OpenAI models. At $2/$6 per million and still untested on the coding-agent index, benchmark it on your own tasks before switching.
Elsewhere: Datasette patches a SQL injection
Datasette 1.0a38, with a 0.65.3 back-port, fixes a SQL injection that let anyone with access to a public table attack instances that mix public and private tables. Update tonight if you self-host it.