OpenAI
OpenAI withdrew three papers from its AI-generated mathematics collection. According to its update log, a sign error in one paper invalidated that proof and two papers built on the same construction. The catalogue went from 722 to 719 papers, and the same update fixed 14 papers and added 6 new Lean
Perplexity
Perplexity has published pplx-decider-v1-27b, a decision model fine-tuned from Qwen3.8-27B, on Hugging Face under Apache 2.0. It returns probabilities for predefined options without writing long replies, and it also powers Perplexity's new Decisions API.
Google made Nano Banana 2.1, its image generation and conversational image editing model, available on October 6, 2026. An update to Nano Banana 2, it improves visual quality, text rendering, character consistency and wide aspect ratios, and is available in the Gemini app, the Gemini API, AI Studio
Perplexity
Perplexity introduced pplx-embed-v2-late, a family of embedding models for searching text and images. The 0.6B and 9B models share one embedding space and can search PDF pages without converting them to text first. The weights are on Hugging Face under the MIT license.
Google DeepMind released EmbeddingGemma 2, which maps text, images, video and audio into one shared vector space. Designed to run on phones and laptops, the model has 740 million parameters in total, and text-only use needs just a 270-million-parameter part. The weights are on Hugging Face and Kaggl
Nous Research
Nous Research, the developer of Hermes Agent, has packaged its open-source agent for companies. Hermes Business gives every team member a hosted agent, a shared balance and shared skills, while Hermes Enterprise runs the same stack on a company's own infrastructure. On October 7 the company said new
OpenAI
OpenAI released 722 mathematics manuscripts on GitHub, produced by an internal model that has not been released yet. According to the company, they are grouped into 372 result families and include claimed solutions to long-standing open problems. Many proofs come with Lean formalizations that a comp
Mistral
Mistral AI has opened a public preview of Mistral Large 4, its largest model to date. The mixture-of-experts model, with over a trillion parameters in total and 52 billion active per step, can be tried through the API on Mistral Studio for now. The company says the weights will be released by the en
GitHub
Microsoft and GitHub announced that the GitHub Copilot coding assistant on Windows can use models running on the PC itself. GitHub's HydraFusion, which routes each task to a suitable model, will now draw on local models alongside cloud ones. Agent commands will run inside isolated environments calle
anthropic
Anthropic has expanded its Cyber Verification Program, which gives vetted security professionals access to its models with fewer restrictions. The program now has three tiers, Defense, Red Team and Specialized, and each includes Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5.
anthropic
Anthropic made Claude Haiku 5.5, the newest member of its smallest model family, available on October 7, 2026. The model is built for high-volume work such as summaries, classification and subagent tasks, and it is the first Haiku with an adjustable effort setting. The same day, Anthropic cut the ca
Google DeepMind has opened SynthID Detector to everyone. The tool finds SynthID watermarks in AI-generated images, video and audio. It works worldwide in English at synthid.com and recognizes content from Google, OpenAI, NVIDIA and Kakao tools, with Apple coming soon.