NPDev Daily AI Briefing — November 15, 2026
No new US frontier checkpoint shipped this weekend — the movement was in open weights, cost curves and corporate risk. Meituan published LongCat-2.5, a 1.6-trillion-parameter agent model trained on Chinese accelerators and released under MIT, while MathArena put Gemini 4 Carbon on top of its math leaderboard at roughly 15x the per-task cost of GPT-6.1 Sol. Nvidia is now run by an interim CEO, and Anthropic prices its IPO on November 19 with its CEO calling compute the company's deepest dilemma.
Written by NPDev Labs with AI assistance and reviewed before publication. How we work.
Nvidia is being run by an interim CEO, and a WSJ report traces $12B in China chip orders to a firm tied to Jensen Huang's former partner
Industry · Confirmed + reported. Nvidia's board announced on November 12 that Jensen Huang resigned as president, CEO and director effective immediately; CFO Colette Kress is interim CEO while keeping the CFO role, Harvey C. Jones is the board's first-ever chair, and the board has begun a search for a permanent successor. On November 15, The Wall Street Journal reported that Megaspeed — controlled since March 2023 by businesswoman Alice Huang, described as Huang's former partner — arranged nearly $12 billion of Nvidia high-performance chip orders for Chinese buyers between March 2024 and June 2025, covering several hundred thousand export-controlled chips. Megaspeed told Reuters that Alice Huang left the company on August 30, 2025, has no current role, and that it conducted lawful business. Why it matters: Nvidia's allocation decisions set the ceiling on every serving stack you rent or buy, and an interim CEO plus a live export-control story is exactly the combination that makes capacity assumptions unreliable. Nvidia has given no cause for the resignation and no change to its roadmap or guidance, and its Q3 results are still scheduled for November 17, so treat supply forecasts as scenario ranges rather than commitments for the next two quarters. Read source: Nvidia newsroom: Reported context: WSJ findings via THE DECODER:
Anthropic prices its IPO on November 19, and its CEO names compute as the company's deepest business risk
Business · Reported. Speaking with Jamie Dimon at a JPMorgan investor event on Friday, November 13, Dario Amodei said "there's a huge uncertainty in what's going to happen next year" and called compute "the deepest business dilemma that the company faces," because capacity must be bought one to two years before anyone knows whether revenue justifies it. Bloomberg-reported commitments total roughly $901 billion, mostly compute and chips, with about $71 billion due by September 2027. The IPO is scheduled for November 19; Anthropic filed its S-1 on October 26 and opened the roadshow on November 10. Why it matters: A $900B commitment book explains a lot of recent product behavior — aggressive pricing on Opus 5.5, the push into agent products, and a $8/$40 per-million price for Fable 5.2. If you are negotiating multi-year inference contracts, treat provider pricing as a function of a company's need to fill committed capacity, and keep a second provider warm rather than assuming today's price holds. Read reported source: THE DECODER: Primary: Anthropic roadshow announcement:
Australia's central bank finds AI hitting young software and customer-service workers first
Labor market · Reported. Reserve Bank of Australia analysis reports that since ChatGPT launched, job postings and employment among 15- to 24-year-olds in AI-exposed fields — software development and customer service named specifically — have fallen more sharply than for older workers in the same occupations. The author, Mike Major, cautions that business-cycle effects and remote-work shifts could also explain part of the gap. Why it matters: This is the first central-bank read on the entry-level rung rather than on aggregate headcount, and it matches what engineering managers report: the tasks that used to be given to juniors (small tickets, test scaffolding, first-pass code review) are the ones models now do adequately. If you run a team, the countermeasure is deliberate ramp work with review, not fewer hires — the senior pipeline is built from exactly those tasks. Read reported source: THE DECODER: Primary: Reserve Bank of Australia:
Meituan publishes LongCat-2.5: a 1.6T-parameter agent model under MIT, trained on Chinese accelerators
Open weights · New today. Meituan's LongCat team released the weights one week after unveiling the model. LongCat-2.5 is a natively multimodal mixture-of-experts model: 1.6 trillion total parameters with roughly 48 billion activated per token, up to 1M-token context and 128K-token outputs, image input tokenized at native resolution rather than bolted on through a separate encoder. The model card lists the license as MIT and credits infrastructure work "on domestic AI accelerators"; THE DECODER reports training ran entirely on Chinese hardware, with the sample deployment on 194 Huawei Ascend chips and no Nvidia GPUs mentioned. Meituan's own table shows 88.4 on Terminal Bench v2.1 against 89.1 for Claude Opus 5, with competitor figures marked as carried over from other vendors' reports. Why it matters: A frontier-adjacent agentic coding model under MIT — commercial use, modification, redistribution — changes the build-versus-buy math for anyone running high-volume coding agents. The benchmark caveat matters as much as the score: the comparison uses a Claude Code harness at 512K context over three runs, so it measures harness plus model, not the checkpoint alone. Downloading 1.6T parameters is not a laptop operation; the practical path is the hosted API or a quantized endpoint. Read source: model card on Hugging Face: Reported context: THE DECODER:
Gemini 4 Carbon takes the MathArena crown — at about 15 times GPT-6.1 Sol's cost per task
Benchmark · Independent. MathArena results reported on November 14 put Gemini 4 Carbon (high) first at 85.4% ±3.0 at roughly $15.70 per task, Claude Fable 5.2 (max) second at 82.8% ±5.3 at roughly $40.20, and GPT-6.1 Sol (max) third at 80.6% ±2.7 for about $1.07. The best open model in the table is Moonshot's Kimi K3.1 Pro at 54.3%. Carbon's most common failure mode is exceeding the time limit, and in one recorded case it left a script running for more than eleven hours without checking on progress. Why it matters: The leaderboard delta is 4.8 points between first and third; the cost delta is roughly 15x. That is a routing problem, not a model-selection problem: keep the expensive model for problems where a wrong answer is worse than a slow one, and route bulk work to the cheap one. The run-away-script behavior is also a harness lesson — put wall-clock and tool-call budgets around any agent, because a model that does not give up will happily spend your budget. Read reported source: THE DECODER: Primary: Google's Gemini 4 Carbon announcement:
Runway opens Praxis-1: Apache 2.0 weights that turn a video model into a robot controller
Open weights · November 11, reported this weekend. Runway released Praxis-1's weights plus inference and fine-tuning code under Apache 2.0, with full fine-tuning and LoRA support, native LeRobot support on Hugging Face, and vLLM-Omni integration. Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together, taking camera frames, text instructions and robot pose as input. Runway reports 46.2% success on NVIDIA's RoboLab-120 across 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5; THE DECODER notes the model occasionally confuses similar-looking objects and cannot reliably plan complex action sequences on its own. Why it matters: Embodied AI just became something a small team can prototype without a vendor contract: 27B parameters runs on a single high-end GPU, and the Apache 2.0 license permits commercial deployment. The published limitations are the useful part — object confusion and weak long-horizon planning mean this is a perception-and-control layer, not an autonomy stack. Pair it with your own planner and safety envelope. Read source: Runway announcement: Research report: Praxis-1 paper:
Practical implications for engineers
Migrate off Claude Sonnet 4.5 before November 30. Anthropic's deprecations page lists claude-sonnet-4-5-20250929 as deprecated, retiring November 30, 2026, with claude-sonnet-5-5 as the recommended replacement; requests to retired models fail rather than degrade. Drop temperature, top_p and top_k from Claude requests. All three were removed in API version 2026-10-21 and now return 400 errors, and earlier API versions reject non-default values on Claude 4.7 and later; the documented replacement is prompting and effort levels, not sampling knobs. Re-check your prompt scaffolding against Opus 5.5 behavior. Anthropic's guidance (published November 12, widely picked up this weekend) says phrases like "think hard" no longer help, that Opus 5.5 always reasons adaptively, and that Medium effort matches Opus 5 at High — while warning that higher effort costs more without reliably improving output. Treat open weights as a cost lever, not just an ideology. LongCat-2.5 (MIT) and Praxis-1 (Apache 2.0) both landed this week with published limitations; the right move is a two-week spike measuring your own eval set against the hosted API before committing hardware or a migration.
Bottom line
The frontier labs shipped no new checkpoint this weekend, and the two things that did move — a 1.6T open-weight agent model from Meituan and a 15x cost spread at the top of the math leaderboard — point the same direction: capability is now roughly table stakes, and the differentiators are cost per task, hardware supply and reliability under long-running agents.
As of
Sunday, November 15, 2026, 15:00 UTC. Window covered: November 13–15, 2026; items dated earlier are labeled with their publication date. Confirmed first-party sources were opened for the Nvidia board action, Anthropic's deprecation and API-parameter pages, the LongCat-2.5 model card, Runway's release post and Google's Gemini 4 Carbon announcement. Benchmark and labor-market items are attributed to their publishers.