Frontier AI, week of 13 November 2026

The busiest model week in a while: Anthropic, Google and Meta all shipped, and almost every launch came wrapped in a benchmark table drawn up by the company selling the model. The quieter stories matter just as much: OpenAI showed how it pulls the plug on its own research runs, Anthropic put a price on its shares, and Australia started talking about prison-grade liability for AI companies. Here is what happened, grouped by theme, with sources.

Models and releases

Anthropic releases Claude Fable 5.2 and Mythos 5.2

On 9 November Anthropic upgraded its two top models, which it says are the same underlying model: Fable 5.2 ships with strong cybersecurity and biology safeguards for everyone, while Mythos 5.2 goes to vetted organisations through its Cyber and Life Sciences Verification Programs. Anthropic says they generate output up to 30% faster than the 5.1 versions and cost less, at $8 per million input tokens and $40 per million output tokens, with an optional fast mode at double that price. In its own table Fable 5.2 scores 69.9% on FrontierSWE v2, against 65.5% for GPT-6 Astra and 61.2% for Gemini 4 Carbon, though Gemini 4 Carbon beats it on AutomationBench (53.9% vs 51.4%). Anthropic also says that in prerelease testing the model formalised the proof of Fermat’s Last Theorem in 30 hours, and that it still recommends Claude Opus 5.5 as the starting point for most coding and knowledge work.

The truth: The most useful line in the launch is Anthropic telling most people not to use it: this is a model for the hardest, longest jobs, priced accordingly. The benchmark lead is real on Anthropic’s chosen tests and gone on at least one of them, and the “same model, two safety levels” split is now how every top lab releases its strongest systems.

Source: Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2

Google opens Gemini 4 Carbon to everyone who pays

On 10 November Google made Gemini 4 Carbon, which it calls its strongest model, available to paying developers and enterprises through the Gemini API, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity. Cyber defenders had it two weeks earlier through Google’s Fairwind Program. Google says it ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at less than half the cost per task of the next best external model. The introductory price of $2 per million input tokens and $10 per million output tokens runs until 30 November, after which it doubles to $4 and $20.

The truth: Read the footnote before you build on it: the headline price lasts less than three weeks. Google also previewed /swarm, which turns one prompt into a team of agents, and that is the real direction of travel: the frontier model as the architect, with cheaper models doing the typing.

Source: Google: Gemini 4 Carbon is now available

Gemini 4 Hydrogen is the cheap, fast model, and it now runs Google Search

Two days later, on 12 November, Google released Gemini 4 Hydrogen for fast everyday coding and agent work, priced flat at $1.50 per million input tokens and $7.50 per million output tokens across a 1-million-token context. Google says it edges past its own bigger Gemini 4 Argon on the Artificial Analysis Intelligence Index (51 vs 50 at the default setting), more than triples Gemini 3.8 Flash’s Terminal-Bench 4.1 score (57% vs 18.2%) and cuts fabricated claims by more than five times. It is rolling out to all users in the Gemini app and in Google Search, which Google calls the biggest model upgrade to Search yet.

The truth: The model in Search matters far more than any benchmark, because that is where billions of people will meet it without choosing it. “Five times fewer fabricated claims” is Google grading its own homework, so the real test is whether Search answers get less wrong in the wild.

Source: Google: Introducing Gemini 4 Hydrogen

Meta’s Muse Ember goes after the cheap end of the market

On 9 November Meta released Muse Ember, its smallest and most efficient model, aimed at the repetitive work agents do all day: tool calls, coding subtasks, routing and document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API, with a 1-million-token context. Meta says it matches Claude Haiku 5.5’s best Artificial Analysis score at about a third of the cost, and its comparison table shows Ember level with Haiku at 36 on the Artificial Analysis Index but behind it on FrontierCode 2.0 (27.2 vs 34.9).

The truth: The race at the bottom of the price list is where most real agent spending happens, so a credible cheap model matters more than another flagship. Meta’s own table includes the test it loses, which is more honesty than usual, and it is still Meta’s chart.

Source: Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember

Research

Claude designed three working gadgets, circuit boards and all

Alongside the Fable launch, Anthropic and Diode Computers, a circuit-board company Anthropic has a commercial relationship with, described asking Claude Fable 5.2 to design a USB speakerphone, a voice-assistant hub and a handheld virtual pet. Anthropic says Claude wrote the specifications, drew the schematics, laid out and routed the circuit boards, designed 3D-printable cases, wrote the firmware and prepared parts lists, and that the prototypes worked without fixes to their circuitry. On Diode’s internal PCB Gym benchmark, Fable 5.2 scored 69% against 25% for Fable 5.

The truth: This is a real step past “AI writes code”: board layout is spatial, physical work. But Anthropic itself says these are untested prototypes with no electrical validation, thermal testing or certification, and human engineers still had to build them, so “AI designs hardware” here means a very capable assistant, not a factory.

Source: Anthropic: Testing Claude on real-world electronics design

Gemini Robotics 2.5 puts a humanoid on a Mercedes-Benz line

On 10 November Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model it tested on more than 200 tasks, such as serving food with tongs and tying a bow. It says success rates rose 30% to 100% over previous baselines, depending on the task, with tasks finished 20% to 30% faster. In a proof-of-concept on Mercedes-Benz production lines in Hungary and Germany, Apptronik’s Apollo 2 humanoid sorted parts into trays with a 96% success rate. DeepMind also released ASIMOV-Jailbreak, an open benchmark of prompts that try to trick robots into unsafe actions.

The truth: 96% sounds great until you remember a factory line runs all day: one miss in 25 picks still needs a human nearby. It’s a pilot, not a deployment, but sorting parts from bins is exactly the dull, repetitive job where humanoids have to prove themselves first.

Source: Google DeepMind: Gemini Robotics 2.5 accelerates real-world general dexterity

Agents and tools

ChatGPT invites businesses to claim profiles and run their own agents

On 10 November OpenAI launched Businesses in ChatGPT in beta. Through a Business Hub, companies can claim and verify their listing, add product feeds with prices and availability, and see aggregated counts of mentions and website clicks; OpenAI says it does not share users’ conversations with businesses. Eligible US e-commerce and online-service firms can also launch a Business Agent that customers can chat with, and businesses can pay to sponsor that agent as a destination for ChatGPT ads. OpenAI says ChatGPT has 1.2 billion weekly users.

The truth: This is ChatGPT turning into a search engine with a business directory and an ad market attached, which is exactly how Google grew rich. The labels separating sponsored agents from ChatGPT’s own answers are the thing to watch, because that line is where trust erodes first.

Source: OpenAI: Introducing Businesses in ChatGPT

Google’s Decisions API answers yes-or-no questions without writing text

On 11 November Google introduced the Gemini Decisions API, powered by a new model called Gemini Decision 1. Instead of generating an answer word by word, it scores a fixed set of options in one pass, handling up to 1,000 yes/no, multiple-choice or rubric questions against one input of text, images, audio, video or PDFs. Google says it returns answers with confidence scores in 150 to 250 milliseconds, at $0.10 per million input tokens with no output charge, and that Google teams already use it to route queries between models.

The truth: Unglamorous and genuinely useful. A lot of agent plumbing is “is this a refund request, yes or no?”, and paying a chatty model to write a paragraph for that was always wasteful. Confidence scores you can route on are the part engineers will actually care about.

Source: Google: Gemini Decisions API: Fast, calibrated multimodal decisions

Policy and safety

OpenAI shows its kill switch for research runs

This week OpenAI described Control Room, a central system that tracks every model workload in its research infrastructure, including training and evaluation runs, and can stop one within 35 seconds of a decision to intervene. Every research job that does agentic inference must be tied to an approved safety case, and the inference engines ask Control Room for permission on every request, so a stopped job simply stops getting model output. OpenAI says the system “fails closed”, denying all requests if Control Room goes down, and that this is live for frontier research workloads and finishing its rollout to the rest.

The truth: This is one of the most concrete safety disclosures of the year, because it describes plumbing rather than principles. Note the limits: 35 seconds is the time after someone decides to stop a job, not the time to notice a problem, and it covers OpenAI’s research systems, not the models it serves to the public.

Source: OpenAI: Control Room

Australia considers a “failure to prevent” crime for AI companies

On 10 November the Australian government said it would consider a new criminal offence for AI companies whose models engage in harmful conduct when the company has not taken reasonable steps to prevent it, the ABC reported. It would also consider criminal penalties for egregious breaches of AI safety standards and an AI safety watchdog able to issue significant fines. The move responds to a review of September’s OpenAI hack of the Medicare statistics portal; the ABC reports that the Australian Federal Police has opened no criminal investigation, and Prime Minister Anthony Albanese called the incident “minor in terms of direct implications”.

The truth: “Considering” is doing a lot of work here: nothing has been drafted, and experts quoted by the ABC expect civil penalties to be more useful in practice. The idea still matters, because “failure to prevent” puts the burden on companies to show they took care, rather than on prosecutors to prove intent.

Source: ABC News: Australia considers world-first criminal charges for AI companies that fail to prevent harm

Industry and compute

Anthropic prices its IPO at $70 to $78 a share

On 10 November Anthropic launched the roadshow for its initial public offering: 825 million Class A shares, almost all sold by the company itself, at an expected $70 to $78 each, with an option for underwriters to buy 123.75 million more. It has applied to list on the Nasdaq under the ticker “ANT”. Reuters, citing the latest draft prospectus, reported that the deal could raise as much as $64.4 billion.

The truth: A raise that size would be enormous by any standard. Last week’s filing showed slowing growth and a huge computing bill, so the price range is really a question to public investors: will you pay for that bill now, on the promise of revenue later?

Source: Anthropic: Anthropic launches roadshow for proposed initial public offering; Reuters via Investing.com: Anthropic IPO to raise as much as $64.4 billion, latest draft prospectus shows

Open source and open weights

Edge0 says it can run a 145-billion-parameter model on a laptop

On 12 November Edge0 published the weights of Edge0-145B and Edge0-120B along with its inference framework. It says the 145B model runs on a compatible laptop such as a MacBook Pro with 29GB peak memory, processes prompts at over 1,100 tokens a second and generates at over 80, using about 120W. On its own evaluations it scores 64.6 on DeepSWE 1.1 and 86.52 on Terminal-Bench 2.1, ahead of the Qwen and DeepSeek models it compares against, and it says the full evaluation setup is being released with its technical report.

The truth: If the memory numbers hold up, that is the story: a model this size on a laptop, offline, with no per-token bill. The benchmark comparisons are the company’s own, and the Claude model it lines up against is Claude Opus 4.6, several generations old, so wait for independent tests before calling it frontier-class.

Source: Edge0: Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer

Baidu previews ERNIE 5.5 Flash, with open weights promised

On 11 November Baidu released ERNIE 5.5 Flash Preview, a mixture-of-experts model with 280 billion total parameters, 14 billion active and a 1-million-token context. It is available now as an API on Baidu’s Qianfan platform, and Baidu says the weights “will be open-sourced under Apache 2.0 soon”. Baidu showcases long agent tasks, including an 11-hour coding run that built a game in Unreal Engine 5.8, and calls the release an intermediate checkpoint whose training is not finished.

The truth: Promised open weights are not open weights until they’re on a download page. Baidu deserves credit for saying plainly that this is an unfinished checkpoint; treat the demo reel as demos, and the Apache 2.0 licence, if it lands, as the real news.

Source: Baidu: Introducing ERNIE 5.5 Flash Preview

Sources were read on 13 November 2026. Spotted an error? That’s the point: tell me.