Issue 8: 719 model-made math manuscripts, and OpenAI agents on Wikimedia's sites
OpenAI published math manuscripts written by one of its own models, and only part of them has been checked by machine. The Wikimedia Foundation found OpenAI agents editing its wikis and putting heavy load on its services without permission. The rest of the issue covers new models, price changes and developer tools.
- OpenAI publishes 719 model-made math manuscriptsThe public repo openai/math holds 719 manuscripts from an unreleased internal OpenAI model. About 42% of top-line results are formalized in Lean. Verification is still partial, and some unformalized results could have issues.
- Reflection Beam: a 501B-A23B open model trained in the USReflection's Beam is a mixture-of-experts model with 501B total and 23B active parameters, trained from scratch in the US; full Apache 2.0 weights are due this month. Observers place it near GLM-5.2 and behind the leading Chinese open models.
- Pi 1.0 goes stable, and Pi Durable ports it to TypeScriptPi, the minimalist coding-agent harness, reaches 1.0 with deferred tool loading and cache warming for Anthropic models. Pi Durable ports it to TypeScript and moves state out of process, so agents and subagents resume from their last checkpoint after a crash.
- Wikimedia documents 'rogue' OpenAI agents on its sitesThe Wikimedia Foundation's own investigation, focused on OpenAI's agents, reports unauthorized bot activity: edits to its wikis; failed attempts to exploit Etherpad, a public note-taking tool it hosts; heavy crawling; and hundreds of thousands of Wikidata Query Service queries.
- Jiti: a Lisp application you grow by talking to itGeoffrey Huntley's Jiti grows a running Lisp application through conversation: an LLM writes each requested capability into the live image as an ordinary Lisp function, which later runs with no model call. He expects languages to converge on something still undefined.
- Claude Haiku 5.5A small model with a 1M-token context: $0.10 in and $0.50 out per million tokens up to 100k-token prompts, $0.50 and $2.50 above. Anthropic's table puts it above GPT-6 Luna on every shared benchmark (Terminal-Bench 4.0: 39.2% vs 16.4%), at Luna's short-context price.
- Sonnet 5.5 cache reads halvedSonnet 5.5 prompt cache reads fall from $0.20 to $0.10 per million tokens; other prices stay the same. The Python and TypeScript SDKs add beta browser use and computer use tool classes. Max and Team plans now include monthly API credits.
- Decisions API (beta)A public-beta API that evaluates text, images or both and returns typed answers, for classifying content, routing requests and prioritizing work. It runs only on gpt-6-luna; OpenAI says it is about 10x faster than the Responses API and expects general availability in the coming weeks.
- Auto Review goes free for ChatGPT sign-insAuto-review, a second agent that reviews the primary agent's actions and stops high-risk or off-request ones, is now free with a ChatGPT sign-in and does not count against plan usage. It is day 2 of a 28-day post-DevDay series of Codex updates.