The Day in AI

• Meituan released LongCat-2.5, a 1.6 trillion parameter model with about 48 billion active per token, and reports 88.4 on Terminal Bench v2.1.
• MIT CSAIL researchers report that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and raised its MATH-500 score from 7.9% to 48.1%.
• Anthropic added message threads in beta to the Claude Messages API, so the server keeps the conversation and later requests send only new messages.
• OpenAI is merging the Chat and Work tabs in the ChatGPT desktop app, and ChatGPT and Codex will share one usage allowance.
• SemiAnalysis launched the GDPx benchmark across 11 agentic domains, and the best model reportedly scores 12%, according to Crypto Briefing.

LongCat-2.5 follows LongCat-2.0 and adds native image understanding. Images are turned into multi-level discrete tokens that sit in the same stream as text, so one model reads both. Training added more long-horizon trajectory data and large-scale reinforcement learning (RL), and Meituan turns failure trajectories from real agents into verifiable training tasks. The context window is 1 million tokens, with output up to 128,000 tokens according to Crypto Briefing.

Meituan reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0. Its AutomationBench score rose from 15.8 to 51.8, and Meituan also reports 91.9 on BrowseComp. All of these are company figures. The model is available through a closed API, and Meituan optimized W4A4 inference (4-bit weights and 4-bit activations) on domestic Chinese chips.

Researchers at MIT CSAIL report that GPT-6 Astra can read raw neural network weights and write edits to them directly. Earlier methods set weights through optimizers such as Adam or Muon, through trained weight-writing networks, or through compilers. In this work the model reasons about the weights in natural language and changes them without gradients or a training loop. The authors say it found the reasoning loop problem that Qwen's model card warns about in Qwen3.5-0.8B and fixed it by editing 0.016% of the weights.

According to the authors, that edit raised MATH-500 from 7.9% to 48.1% under the same sampling settings, and IFEval improved about threefold. In a second test on the Kinetix car-on-ramp task, changing 3.5% of a policy's weights took success from 0% to 100% over 1,024 episodes. The Atlas could only partly confirm the figures, because some were not in the visible page text. It is unknown whether the method works on larger models.

Anthropic's message threads, in beta, store a conversation on the server so each later request carries only the new messages. Smaller requests give long conversations and agent loops a lower time to first token. A thread expires 5 minutes after its last request and is deleted no more than an hour after its last use.

• OpenAI is folding ChatGPT's Chat and Work tabs into one experience for files, plugins and web actions, and Codex stays separate in the desktop app.
• DeepSeek added a creator mode to DeepSeek Harness that builds plugins such as a 3D model renderer from a description, two days after releasing Super-agent.
• Tencent Cloud launched WAND-Vega, a multimodal model for images, video and speech, on 30 October, and says WAND serves more than 2,000 companies.
• ServiceNow released Apriel Voice 27B Preview on 29 October, a speech-language model built on Gemma 4 26B-A4B that handles speech and tool calls in one model.
• Apple says Apple Intelligence in China, using Alibaba's Qwen, has regulatory approval, with a wider launch reported for December in iOS 27.3.
• NVIDIA is reportedly facing Justice Department questions over its $12.93 billion Hugging Face deal, according to Startup Fortune.
• OpenAI has GPT-6.2 Astra appearing in Codex usage ahead of a possible launch, according to TestingCatalog.
• Anthropic confirmed new Claude Code limits that steer it away from recreating specific existing games, according to Tom's Hardware.
• Tutor, which builds warehouse robots and a robot foundation model, raised a $75 million Series B led by a16z, according to Forbes.
• Terence Tao published an essay titled "Mathematics has changed. Let's act like it."