GPT-5.6 splits in two, the free tier gets ads, and two dates to know

OpenAI updates from the past month: a model change, a bigger and stranger free tier, two dates, and a math result worth a raised eyebrow.

1. 01The ChatGPT model line changed on August 6↗. Paid users get an updated GPT-5.6 Sol with an effort slider that trades speed against depth per response, and free accounts move to GPT-5.6 Luna with a Think button for harder questions. When the model under a workflow changes, this department re-runs the workflow against its evaluation set before trusting it again. Your personal prompts deserve the budget version of that habit: spot check the outputs you rely on before you rely on them again. And notice what the effort slider actually is: the same speed-versus-depth tradeoff we make when routing jobs between frontier and local models, now sold as a consumer feature.
2. 02The free tier got bigger, and it got ads. Free and Go accounts gained unlimited text chats the week of August 10↗, and ads, which have run on the United States free tier since February, expanded on August 11↗ to the United Kingdom, Mexico, Brazil, Japan, and South Korea. Paid and Enterprise tiers stay ad free, and OpenAI says ads do not influence answers and conversations stay private from advertisers. The detail that matters most for work: free accounts are enrolled in model training by default. A rate sheet pasted into a personal free account is a rate sheet volunteered as training data. See the callout below.
3. 03The one that already hit its deadline: ChatGPT Enterprise retired individually authorized sync↗ for connected apps such as Google Drive. New individual connections stopped August 10, existing ones shut off August 14, and access moves to admin managed sync. If a ChatGPT workflow of yours reads from a connected app, confirm it still works and flag it to IT if it does not. The direction matters as much as the date: access to business data is moving from individual grants to admin managed provisioning across the industry, and the Claude Corner in this issue reports the same shift. That is good news for anyone who ever has to answer an audit.
4. 04A second date: the official DALL·E GPT retires from ChatGPT on August 30↗. Download any images you want to keep before then. Image generation itself is not going anywhere: ChatGPT Images is the replacement, and user-created GPTs with image generation enabled are unaffected. If you built anything on the DALL·E GPT specifically, migrate it this month.
5. 05OpenAI announced that an internal, unreleased model called Astra solved ten open problems in mathematics and theoretical computer science↗, publishing machine checkable proofs, for roughly 2,000 dollars in compute. The results are still being independently examined, and skepticism is the right posture until they are. But hold the cost figure up to the light: problems no human had cracked, for the price of a crane tire. When the ceiling of what models can do moves like that, the list of Bigge tasks once ruled too hard for automation deserves a re-read. We keep that list, and announcements like this are why it gets re-read.

Same routing as always. ChatGPT license and training questions go to the Innovation team, and build ideas come here.