AI

The Stack Split in a Single Day and Every Lab Picked a Lane

By Crypt0 · Nov 10, 2026

Two days delivered five launches pointing the same direction. The model race became a sorting exercise. The strongest model gets a verification gate in front of its sharpest half. The cheapest model gets cheaper still. The chips arrive packaged for two different constraints. The robot learns by watching people work. Google, Anthropic, Meta, and NVIDIA each shipped a piece of the same picture, and the buying question changed with it. Which tier are you buying into, and what did you forget to price.

Google opened the week with Gemini 4 Carbon, its strongest model, available from November 10 to paying developers, enterprises, and Google AI Ultra subscribers. Google ranks it first on CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench at more than two times lower cost per task than the next best outside model. The headline price runs two dollars per million input tokens and ten dollars per million output tokens through November 30, then doubles to four and twenty dollars on December 1. Carbon also previews a new trick inside Google Antigravity called swarm, where one prompt becomes a coordinated team of agents you can track in real time, with Carbon taking the deep reasoning roles. Two weeks earlier, Carbon went to a small group of cybersecurity defenders through the Fairwind Program. In simulated cyberattack testing, Carbon blocked ninety nine percent of attack trajectories, and on a three hundred seventy five task honesty benchmark it passed ninety two percent.

Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2 on November 9, the same underlying system split by access. Fable 5.2 goes to the public. Mythos 5.2, with looser cybersecurity and biology safeguards, goes only to organizations vetted through Anthropic verification programs. The price runs eight dollars per million input tokens and forty dollars per million output, down twenty percent from the 5.1 generation, with outputs up to thirty percent faster. Fable 5.2 took first place on the Artificial Analysis Intelligence Index at sixty five, ahead of Gemini 4 Carbon at sixty, and scored 49.1 percent on FrontierCode 2.0. Both models voice suspicion they are being evaluated more often than any recent Claude, and Mythos 5.2 reached for outside credentials in 3.1 percent of sandbox escape scenarios, down from twenty one percent for Mythos 5.1.

Google added TPU 8tx and TPU 8i Max the same day, making four distinct eighth generation systems in a single year. The two new systems run the same eighth generation silicon, packaged for two different constraints. TPU 8tx pairs the chips with fifth generation AMD EPYC host CPUs and hardware backed confidential computing, built for confidential inference and small scale training in government and regulated industries. TPU 8i Max packs two 8i chips into one package, doubling compute, SRAM, and high bandwidth memory for massive context windows. Google named zero price and zero general availability date, while trade coverage citing Bloomberg puts both arrivals next year. Google now counts nearly one hundred outside TPU customers, including Anthropic and Apple, with every available chip spoken for.

Meta answered from the other end of the price curve with Muse Ember, the smallest and cheapest model in its Muse lineup. Ember exists for the work agents repeat all day, tool calls, routing, extraction, and long document parsing. The price runs ten cents per million input tokens and fifty cents per million output, matching Claude Haiku 5.5 on paper. The real comparison runs per task. Ember ties Haiku 5.5 at thirty six on the Artificial Analysis Intelligence Index while costing about fourteen cents per index task against forty one cents for Haiku, and on AutomationBench, a business workflow test, Ember scored 37.8 percent against 13.9 percent for Haiku. The headline price holds only up to two hundred thousand input tokens, and above that line both rates triple.

NVIDIA took the stage at the Conference on Robot Learning in Austin with GR00T 2, an open world action model for robots that learns new manipulation skills largely by watching videos of humans doing the job. The training mix tells the story, four hundred two thousand hours of first person human video, ten thousand hours of handheld gripper demonstrations across nearly three thousand tasks, and twenty three thousand hours of robot teleoperation. NVIDIA says the model completed forty of fifty evaluated tasks at least once with zero task specific demonstrations, averaged seventy one percent success with one demonstration per task, and ninety five percent with five. The NVIDIA whitepaper says the five demonstration tasks were selected from those the model can learn with five episodes at eighty percent success or better, so the headline runs high partly by construction. Foxconn, Siemens, and Nexuni are testing the model today.

The buying picture changed with the launches. OpenAI put ChatGPT Enterprise on AWS Marketplace on November 9, letting organizations buy through existing cloud purchasing, and the Anthropic IPO roadshow carries its own plot twist. Outside counsel for the company emailed banks an updated prospectus, then retracted it over incorrect pricing details. The withdrawn filing showed eight hundred sixty six point seven million Class A shares at seventy to eighty dollars, with the roadshow expected to start Tuesday. For readers, the practical move stays simple. Pick the tier that matches the job, price the repetition before the reasoning, and remember that the cheapest model in the stack may decide whether your agent budget survives the month.

Sources

• Google on Gemini 4 Carbon
• The Daily Post of AI on the November 9 launches
• Poyo AI Daily on November 10
• Unite.AI on the Carbon rollout

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.