OpenAI sets off a math verification race

OpenAI’s latest math release has researchers racing to understand proofs that arrived faster than human explanation.

Today’s issue covers AI-made proofs, Claude’s cheaper small model, and a safe starter workflow for agents.

OpenAI released a large batch of AI-generated mathematics work this week, and researchers immediately moved into verification mode.

Computer scientist Scott Aaronson called it one of the biggest days in mathematical history, while also making the hard part clear: people still have to understand what the model produced.

The release included 372 claimed major results, including a proof of the Unique Games Conjecture, a problem that has shaped theoretical computer science for years.

Some proofs reportedly have Lean certificates, which means formal software can check parts of the logic. That still does not make the work easy for humans to read, explain or trust.

The hard part is the workflow around the breakthrough.

Aaronson also contrasted OpenAI’s public release with Anthropic’s quieter route, where human mathematicians helped write up a separate AI-assisted result before release.

Founders context: If your team uses AI for legal, finance, security or engineering analysis, separate generation from acceptance. Name the reviewer, keep the review trail, define the acceptance test, and make one person responsible for signing off the result.

Anthropic launched Claude Haiku 5.5 on the 7th of October with a pricing model built for heavier usage.

Haiku is the small Claude model built for high-volume tasks: summaries, classification, routing, compaction, live support and other jobs where speed and cost matter more than heavyweight reasoning.

Anthropic says Haiku 5.5 is about 75% cheaper to run than Haiku 4.5. It also says Sonnet 5.5 cache reads are being halved, making Sonnet around 20% cheaper on most agentic work.

Then comes the subscription push. Claude Max and Team users now get monthly API credits:

Anthropic is pushing customers toward products and internal agents built on its platform, not only chats inside Claude’s web app.

For teams, the useful test is where each job should run. Send repetitive work to the cheaper model, keep harder judgement calls on the stronger model, and treat usage credits as a discount, not cash.

Founders context: For any recurring AI workflow, price three lanes separately: cheap routing, mid-tier drafting or summaries, and expensive final judgement. If the numbers only work when every step uses the best model, the workflow probably is not ready to scale.

A practical agent setup does not need to start with the agent reading your whole inbox.

Allie K. Miller, who trains teams on AI adoption, shared a cleaner pattern this week: give the agent its own email address first, then route low-risk inbound mail to that address.

The first use case is junk signup filtering. Instead of handing an agent access to the main inbox, use the agent address for newsletters, trials, downloads and low-value forms.

The agent reads that stream and sends back the useful pieces as recaps. The human keeps the main inbox clean and only sees what earned attention.

The side inbox also works for other low-risk streams.

The setup limits risk because the agent has a narrow inbox, a narrow job and no automatic path into sensitive systems.

Founders context: A useful first test is measurable: send 50 low-risk messages to the side inbox, then check how many the agent sorted correctly, how many it escalated, and how much cleanup it created. Broader access comes only after that pass.

🖥 OpenAI started rolling out Intelligent UI in ChatGPT. Ethan Mollick, a Wharton professor who studies AI at work, said early access felt like a move away from walls of text toward interfaces generated for the task at hand.

🧾 Claude API credits now come with Max and Team plans. The Help Center lists $100 for Max 5x, $200 for Max 20x, and pooled Team credits capped at $500.

🎼 Simon Willison found Claude Opus 5.5 producing better game music than expected. Willison is a long-time developer and AI tooling writer, so his quick experiments are useful early product signals.

📐 Virginia Williams and Josh Alman posted an AI-assisted algorithms result. Aaronson’s write-up says Anthropic helped on the breakthrough, while the human authors released the digested paper.