GPT-6 reaches free ChatGPT; OpenAI withdraws three math papers

GPT-6 Sol and GPT-6 Luna, which OpenAI released in its API on September 22 but kept out of chat, are now in ChatGPT. Plus, Pro, Business and Enterprise subscribers get Sol from October 7, and free and Go users move from GPT-5.6 Luna to GPT-6 Luna in chat from October 8. Both arrive with a feature OpenAI calls Intelligent UI: the model decides when an answer should include an interactive diagram, chart, form or calculator, and builds it inside the reply. Users can turn the visuals down if they prefer text. OpenAI also says the models search the web better and resist attempts to get around their safety training more strongly.

Anthropic shipped Claude Haiku 5.5 on October 7, the small model it had promised in late September. For prompts up to 100,000 tokens it costs 10 cents per million input tokens and 50 cents per million output tokens, against $1 and $5 for Haiku 4.5, and Anthropic puts the average saving at about 75 percent. It is the first Haiku with an adjustable effort setting, and Anthropic pitches it as a fast subagent for summaries, compaction and browser use rather than for complex coding. The same day, Anthropic halved Sonnet 5.5's cache-read price and announced monthly API credits for Max and Team subscribers.

A day after publishing 722 math papers produced with an unreleased model, OpenAI has withdrawn three of them. A sign error in a paper on Weil classes broke an argument that two other papers depended on, including one claiming the rational Hodge conjecture for products of K3 surfaces. OpenAI also repaired proofs in 14 other papers and says about 42 percent of the remaining results, 300 of 719, now have machine-checked Lean proofs. The rest rely on written arguments, which is where a sign error can hide.

In a three-month randomized trial at eleven law firms, economist David Autor and Google researchers gave 133 patent lawyers an AI drafting assistant or held it back. With the tool, drafts improved for everyone. The test that mattered came at 90 days, when every lawyer had to correct a flawed patent without AI. Senior lawyers who had used the tool outperformed their peers, which suggests they learned from it. Junior lawyers showed no average gain: their scores split, with more strong results and more very weak ones. The paper is out through the National Bureau of Economic Research.

At its Windows and Surface event in San Francisco, Microsoft showed Copilot finding tax documents across folders on a PC, renaming and zipping them, and drafting an email to an accountant. Microsoft calls the approach Hybrid Intelligence because it mixes models on the device with models in the cloud, and says it reaches Copilot over the next couple of months. Windows 11 also gains Execution Containers, a way to sandbox AI agents, which Satya Nadella said will reach every Windows 11 user. The new Surface Laptop Ultra, built on Nvidia's RTX Spark chip to run models locally, starts at $2,599.

Biohub, the research nonprofit founded by Mark Zuckerberg and Priscilla Chan, says partners have committed $1.8 billion in funding, data and computing to generate biological data for AI models that predict how cells behave. Google DeepMind, Meta and Isomorphic Labs are putting in $300 million together. The US Department of Energy will put in more than $500 million over five years, and the National Institutes of Health is contributing existing datasets. The aim is open data that lets researchers run some experiments digitally before running them in a lab.

Google has opened a website, SynthID.com, where anyone can check whether an image, video or audio file carries a SynthID watermark. Until now, outsiders had to ask Gemini. The checker now reads the watermarks of Google's partners as well, including OpenAI, Nvidia and Kakao, with Apple to follow. It answers yes or no rather than highlighting the marked regions, needs a Google, OpenAI or Apple sign-in, and allows about ten checks a day. Google says Gemini alone has watermarked more than 180 billion images and videos.

Michael Smith of North Carolina was sentenced to 18 months in prison and ordered to forfeit more than $8 million, after pleading guilty to the first US criminal case of AI-assisted streaming fraud. Over seven years he made hundreds of thousands of AI songs and ran thousands of fake accounts on Spotify, Apple Music, Amazon Music and YouTube Music to stream them. In one month of 2023 his tracks drew 80.9 million streams, against 9.3 million for Taylor Swift's whole catalog. Prosecutors say the payouts came out of the royalty pool shared with real artists.