OpenAI published 722 math manuscripts from an unreleased model, and launched GPT-6 in ChatGPT
OpenAI published 722 mathematics manuscripts, grouped into 372 families of related results, all produced by an internal frontier model it has not released.
OpenAI put the manuscripts in a public GitHub repository on 6 October. Each family holds a main result plus supporting arguments or alternative proofs, and many of the proofs come with Lean formalizations, so a proof checker can verify those steps mechanically. OpenAI says more formal proofs will be added. It follows the company's earlier Navier-Stokes announcement.
OpenAI estimates that the average result used about three hours of ChatGPT Pro thinking compute. It also released 10 detailed summaries of the model's reasoning and statistics on attempted problems, which Latent Space and other outlets put at about 4,000. Ynetnews reports that some results relate to three of the Millennium Prize Problems. OpenAI consulted the IAS Advisory Group on Mathematics and AI on how to release the batch and says the results are at different stages of verification and may contain errors. How many of the 372 families hold up under outside review is not yet known.
On 7 October OpenAI rolled out GPT-6 in ChatGPT with a feature it calls Intelligent UI. Answers can now mix text with charts, diagrams, forms, tappable buttons and small tools such as a calculator, and the model picks the format. OpenAI built a library of streamable components and a compiler that draws the interface while the model is still generating. Paid tiers run on Sol and free users on Luna, and OpenAI says the model is built for ChatGPT's more than 1.2 billion weekly users. A day earlier OpenAI opened the Decisions API in public beta on gpt-6-luna. It returns a probability, a choice from a fixed set or a rubric score for text and images, costs $0.10 per million input tokens with no output charges, and OpenAI says it runs about 10 times faster than the Responses API.
Anthropic released Claude Haiku 5.5 on 7 October, its smallest and, by its account, fastest model. It has a 1M token context window and is the first Haiku with adaptive thinking, where the model decides how much to reason within an effort setting the developer chooses. Requests up to 100K tokens cost $0.10 per million input tokens and $0.50 per million output tokens, and longer ones cost $0.50 and $2.50. Anthropic reports 39.2% on Terminal-Bench 4.0, where Haiku 4.5 scored 0.0%. It also halved cache read prices for Sonnet 5.5, which launched on 28 September.
Also in the news
- Rohit Prasad, who led Amazon's AGI organization and the Nova models, will become CEO of Boston Dynamics, the Hyundai Motor Group robotics company.
- NVIDIA and Microsoft announced RTX Spark, an Arm-based Windows platform with up to 128GB of unified memory, and the Surface Laptop Ultra on it starts at $2,599 and ships 16 October.
- Google opened a public SynthID Detector site that checks images, video and audio for watermarks from Google, OpenAI, Nvidia and Kakao, with about 10 checks per user per day.
- Google DeepMind released EmbeddingGemma 2, an open 740M parameter model that puts text, code, images, video and audio into one 768-dimension embedding space.
- Google DeepMind released Nano Banana 2.1, an image generation and editing model with better mask-based editing, listed on OpenRouter at $1.50 per million input tokens.
- Nous Research reportedly raised a $90 million Series B at a $1.5 billion valuation to scale its Hermes Agent.
- xAI founder Elon Musk said Grok Bot will route some tasks to rival models including Claude Opus 5.5, according to The Information.
- Vinci, an engineering AI startup, reportedly raised $250 million at a $1.5 billion valuation.
- Nvidia published a write-up on fine-tuning one Nemotron model family to gold-level results at both IOI and IMO.