The Day in AI

• OpenAI released Images 2.6, which makes a GIF from a prompt or photo in one step and is rolling out to all ChatGPT and Codex users.
• Anthropic released FlashProve on 2 November, an open-source prototype that proves a model response came from a committed set of hidden weights.
• Cat Wu, head of product for Claude Code at Anthropic, announced on 3 November that she has left the company.
• Meta acquired Conductor, a tool for running several AI coding agents in parallel, according to Axios.
• DeepSeek reportedly plans to train its V5 model on Huawei Ascend chips, according to TechNode.

ChatGPT can now turn a text prompt or an uploaded photo into a GIF in one step, which it could not do natively before. Users can change the playback speed, edit single frames and share the result. Images 2.6 also generates still images faster, and OpenAI is launching GPT-Image-2.6 Flare and Sunburst in its API. OpenAI reports over 4 billion images generated per week across ChatGPT and the API, up from the 3 billion a week it reported in September.

Anthropic released FlashProve on 2 November as an open-source research prototype. It lets a provider show that a response came from one exact set of model weights without revealing those weights. The provider publishes a short commitment to the weights, and each prompt and response pair gets a zkSNARK proof, a compact cryptographic proof that the computation ran as claimed, which anyone can check on an ordinary CPU. The design is software only and does not need trusted hardware.

Anthropic demonstrated it on Llama 4 Maverick, a mixture-of-experts model. The prover only does work for the roughly 17 billion parameters active per token, and the choice of experts stays hidden. Anthropic reports proving at 110 tokens per second, with each proof at 18.78 MB and verification taking about 17 seconds on 24 CPU threads. Proving is still expensive. By Anthropic's measurement it takes about 72 times the per-GPU time of serving the same model with FP8 in vLLM on eight H100s. Whether any provider will run it in production is not known.

Cat Wu led the product side of Claude Code alongside its creator, Boris Cherny, and also oversaw Cowork. She announced on X on 3 November that she had left Anthropic the previous Friday, 30 October, and Anthropic confirmed it. Wu hired some of Anthropic's product managers. Her next move and her reason for leaving were not stated.

• Meta acquired Conductor, and CEO Charlie Holtz says the team will join Meta Superintelligence Labs, according to Axios.
• DeepSeek reportedly plans to train V5 on Huawei Ascend chips, according to TechNode.
• Pangram released Pangram Image 1.0 for detecting AI-generated images and video, and says its lighter Flash model reaches 98.57% accuracy on AI videos.
• Apple says Apple Intelligence and Alibaba's Qwen have regulatory approval in China, and a wider launch is reported for December in iOS 27.3.
• Sarvam AI released Prism on 3 November, a decision model built on Sarvam 105B that Sarvam says averages 71.8 across three test sets.
• Baidu launched Apollo World on 3 November, a driving simulator that Baidu says produces up to 100,000 kilometres of simulated driving a day for Apollo Go.
• OpenAI described Pando on 2 November, an internal Git backend built with Codex in 10 weeks that OpenAI says cut push latency by almost 80%.
• Sentra raised $150 million to secure autonomous AI agents, according to the Wall Street Journal.
• Shakeel Hashim of Transformer argues the Democratic sweep of Congress makes frontier AI legislation with mandatory testing and developer liability more likely in early 2027.