AI News | November 3, 2026

Lead: In the past 24 hours, AI news followed one thread: models are starting to do the work themselves, while rules are starting to draw lines around them. Meituan released a 1.6-trillion-parameter native multimodal model, DeepSeek open-sourced a super-agent that listens to repositories, and robots got a simulator to practice in before touching real hardware. Meanwhile Anthropic told Claude Code not to recreate existing games, and Apple confirmed Qwen will power Apple Intelligence on iPhones in China in December — capability boundaries and usage boundaries redrawn on the same day.

1. Meituan's LongCat-2.5: 1.6T parameters, ~48B activated per token, built for long-horizon agent tasks

On November 2, Meituan's LongCat team released LongCat-2.5, a native multimodal model with 1.6 trillion total parameters and about 48 billion activated per token — a sparse mixture-of-experts design. It builds on LongCat-2.0 with stronger native visual understanding and agentic coding, and is positioned for long-horizon agent tasks. "Native multimodal" means text and images flow through one model rather than through an external vision module.

Why it matters: The headline number is less important than the arithmetic: activating only 48B at inference time is what makes a trillion-parameter model affordable for long tasks. The positioning word to watch is "long-horizon" — the contest is shifting from single-turn Q&A to whether a model can decompose a task, execute it over many steps, and stay on track. Two caveats: official benchmarks have yet to be matched by independent evaluation, and weight/API availability is still unannounced — so for now this is a Chinese foundation model worth tracking, not yet one to build on.

Sources: IT Home (Nov 2, 2026), via https://aigc.news/events/longcat-2-5-ai-1-6t-2026-11-02-143dd349/ · iiMedia, Nov 2 new-product roundup https://www.iimedia.cn/c1088/119408.html

2. DeepSeek open-sources dsh-superagent: a router-worker architecture that splits repo tasks into parallel tickets

DeepSeek's Harness team open-sourced dsh-superagent v0.1.0 on November 1 under the MIT license, a research prototype for next-generation autonomous agents. Its router-worker architecture listens to GitHub (optionally Feishu) task requests, breaks repository work into parallelizable tickets, and dispatches them to multiple agents running in their own sessions and directories. The project page stresses this is a research prototype, not an official product or production-ready software.

Why it matters: The interesting shift here is organizational, not just capability: the era of one agent doing one thing at a time is giving way to "one dispatcher plus many workers." Developers should try it in an isolated repository, on low-risk tasks, with narrow permissions and human approval before any merge — those three guardrails are non-negotiable for a prototype. Its real value is letting the community test DeepSeek's answer to the super-agent question early, before committing their own agent architecture in that direction.

Sources: MarkTechPost coverage, via Communeify AI Daily Nov 2 https://www.communeify.com/en/blog/ai-daily-2026-11-02/ · Enterprise DNA: DeepSeek publishes super-agent prototype (Nov 1) https://enterprisedna.co/resources/ai-pulse/ai-pulse-2026-11-01-deepseek-publishes-super-agent-prototype/

3. UBTECH releases Thinker-Sim: an embodied-intelligence simulator so robots practice in software before real hardware

On November 2, UBTECH announced Thinker-Sim, an embodied-intelligence simulation platform built on its Real2Sim2Real framework. It does three things: bring real scenes into the computer, generate training data in bulk, and score robot capabilities — so skills are practiced in simulation, where failure is cheap, before being transferred to real machines.

Why it matters: Data, not models, is the most expensive part of embodied intelligence: collecting it on real robots costs time, space, supervision, and occasionally hardware. A simulator moves "practice" into a reproducible, batchable, scoreable virtual world. Read alongside the November 1 report of Meta and Fourier building a humanoid prototype in three months, the field is filling in both halves of the puzzle — the body and the training ground. The hard part remains the sim-to-real gap: skills learned in simulation must survive friction, lighting, and sensor noise on real machines, which only transfer tests — not launch copy — can prove.

Sources: iiMedia, Nov 2 new-product roundup https://www.iimedia.cn/c1088/119408.html · Physix Frontier Headlines, Nov 2, 2026 https://daily.physixfrontier.com/issues/2026-11-02.html

4. Anthropic draws a line in Claude Code: no recreating existing games, even renamed or called original

Anthropic has updated Claude Code's system prompt to stop it being used to recreate existing video games, reported The Verge, Polygon, and others on November 2. The two new sentences are blunt: "Do build original games, games in a genre, and games from the user's own designs and assets. Don't build a recreation of a specific existing game, even if it is renamed or called original." Technical staff member Addy Osmani confirmed the change, citing Anthropic's usage policy against IP infringement; original games, genre games, and work from the user's own assets are meant to keep working. The rule ships as a server-side instruction (internally "tengu_heron_brook") that users cannot disable, against a backdrop of AI clone games flooding Steam Next Fest in October.

Why it matters: This is platform governance arriving as a two-sentence prompt patch — no press release, no policy paper. For game developers the line is reasonably clear: copying a specific game is out, working in a genre or from your own designs is in. But grey zones are real: a system prompt is an instruction, not a filter, and can be talked around; the model also cannot reliably verify who owns a design. There are already reports of original projects being refused, and Anthropic says it will keep iterating on the balance. The general lesson for anyone building on AI: upstream compliance boundaries can arrive as silent updates, and any workflow that assumes "the model will do anything" can be rewritten by a single prompt change.

Sources: Polygon (Nov 2) https://www.polygon.com/anthropic-claude-ai-game-replicas/ · PC World (Nov 2) https://www.pcworld.com/article/3274819/claude-wont-reverse-engineer-games-for-you-anymore.html

5. Apple confirms: Apple Intelligence lands in China in December, powered by Alibaba's Qwen

Alibaba announced on October 31, as relayed by RTHK and multiple outlets on November 2, that its Qwen models will power Apple Intelligence in China. The first features arrive with iOS 27.3 in December, with developer-beta testing from November 4. Qwen will plug into Siri, system-wide Writing Tools, and Visual Intelligence as an option — the first time Apple has handed system-level AI in China to a local model.

Why it matters: For iPhone users in China, this is one of the year's most concrete AI landings: Siri stops being just a chattier voice assistant and gains a local foundation model underneath, while Writing Tools and Visual Intelligence put system AI into everyday typing and photography. Note the word "option": Apple is not hard-wiring Qwen as the only path, keeping room to balance compliance and experience. Combined with the November 1 news of Qwen AI glasses gaining translation features via OTA, Alibaba's personal-agent push is advancing on phone and glasses at once — and the contest is shifting from parameter counts to who sits closest to the user.

Sources: AIdapted, AI News of the Day – November 2, 2026 (per RTHK) https://aidapted.ro/en/articles/ai-news-november-2-2026-apple-qwen-deepseek-pagebreak/ · Communeify AI Daily, Nov 2 (Chinese edition) https://www.communeify.com/cn/blog/ai-daily-2026-11-02/

6. OpenAI merges ChatGPT's Chat and Work tabs: from asking to getting it done without switching, Codex stays separate

Per TechCrunch's November 2 report on an email OpenAI sent users, the ChatGPT desktop app will fold its Chat and Work tabs into one experience — first to a test group for feedback, then widely. Users will move from exploring an idea to having ChatGPT carry it through in one place, bringing in files and plugins to build a website, create a presentation, or act on the web while reviewing its work; Codex remains a separate experience in the desktop app. In the same batch of product changes, ChatGPT's web, iOS, and Android clients gained account switching, with up to two accounts on one device (help docs updated October 31).

Why it matters: Read this together with yesterday's news that ChatGPT and Codex will share one token allowance from December: OpenAI is fusing the Q&A tool and the working agent into a single product, on interface and billing alike. For users, losing the tab switch is real convenience — no more copying a thread from Chat to Work to restart a job. But allowance management matters more: heavy Codex afternoons can now eat into evening ChatGPT budgets. Anthropic merged Claude's chat and Cowork into one screen in September; both leading labs making the same call in one quarter confirms "conversation as workbench" as the new consensus — the next contest is whose agent actually finishes the job.

Sources: TechCrunch (Nov 2) https://techcrunch.com/2026/11/02/openai-is-finally-merging-chatgpts-chat-and-work-tabs/ · iiMedia, Nov 2 new-product roundup (account switching) https://www.iimedia.cn/c1088/119408.html

7. a16z's "rack spread": token prices down 67% in six months, GPU rents up over 100% — inference math gets harder

In its Charts of the Week published October 31, a16z flagged a counterintuitive combination: average token prices have fallen 67% since May 2026, yet total token spending rose 122% to 160%. Over the same stretch, Nvidia B300 rental prices rose 106% since mid-July, B200 rose 66%, and H100 rose 17%. a16z calls the gap between ever-cheaper tokens sold and ever-pricier GPUs rented the "rack spread," explained by the Jevons paradox: cheaper inference drives usage up faster than prices fall, keeping hardware scarce.

Why it matters: This punctures a common illusion — "falling token prices mean falling AI costs." For inference providers, revenue is priced per token while costs are rented per GPU-hour; if those curves move in opposite directions, margins depend entirely on throughput gains outrunning rent increases — sell more and you can lose more. For buyers, the lesson flips: watch the total bill, not the price list, because usage growth usually outruns per-unit declines. For the industry, the spread is a leading indicator of the compute cycle: if rents keep rising alongside usage, scarcity is structural, not a blip; if usage plateaus, high rents could reverse. a16z itself cautions the data covers only part of demand — directional, not precise.

Sources: Crypto Briefing (Nov 1) https://cryptobriefing.com/a16z-ai-token-spending-surges-prices-fall/ · iaipie, Nov 2 AI industry roundup https://iaipie.com/2026%E5%B9%B411%E6%9C%882%E6%97%A5ai%E8%A1%8C%E4%B8%9A%E8%B5%84%E8%AE%AF%E9%80%9F%E8%A7%88/

Closing: When models start breaking down tasks themselves, practicing in simulation, and moving into phone operating systems, the contest shifts from "whose model is bigger" to "whose agent gets work done, whose rules hold, and whose books balance" — and Meituan's 1.6 trillion parameters are a reminder that the foundation-model race is far from over; its battlefield has simply moved into tasks, simulators, and bills.

Chinese version: AI 大事件 | 2026年11月3日.