Three threads run through today’s issue.
The first is who grades the labs. OpenAI rated its own GPT-6 Cyber Critical for risk after it found 81 new Firefox bugs, METR says Anthropic’s models are improving about twice as fast as before, and Dwarkesh Patel’s bluffing game put GPT-6 Astra ahead of Opus 4.6.
The second is handing AI the keys. Anthropic now lets Claude read US subscribers’ bank accounts, while Cloudflare has started a coalition because sites can no longer tell a person from the AI acting for them.
The third is price. Google’s Gemini Omni 1.2 Flash got 30% cheaper, Devin Review now costs 25 to 75 cents a pull request, and Thinking Machines’ Inkling 2 Small uses 17B of its 594B parameters per token.
The Report Card: Labs Graded by Themselves and Everyone Else
Three very different tests landed on the same two companies this week. One was written by OpenAI about its own model, one by an outside evaluator, and one by a podcaster with a social deception game.
- OpenAI launches GPT-6 Cyber and rates its own model Critical for risk. GPT-6 Cyber scored 48.5% on an exploit test, up from 32.1% for its predecessor, and found 81 new high-severity Firefox bugs. Access is limited to approved customers, and OpenAI’s own framework rates it Critical for cybersecurity.
- Outside group finds Anthropic’s AI is improving twice as fast as before. After eleven weeks inside Anthropic, METR estimates benchmark scores have climbed 1.9 times faster since October 2025, with a wide range of 1.5 to 2.8. Its best guess is that AI doing the research explains most of the speedup, though it calls the evidence soft.
- Dwarkesh Patel made AI models lie to each other, and GPT-6 Astra won. After 386 games of Blood on the Clocktower, GPT-6 Astra rated about 430 Elo points above Claude Opus 4.6, which Patel says means it would win roughly 92 percent of matchups. Patel’s reading is that OpenAI’s models weigh each claim against several days of events, while Anthropic’s tend to take reported information at face value.
Who Is Really Asking: AI With Access to Your Life
Software that acts for you needs your accounts and your credentials. Two announcements show both sides of that bargain.
- Claude can now read your bank accounts and answer questions about your money. Anthropic has started a US beta of Personal Finance for Pro, Max, and Ultra subscribers, with read-only access to bank and card accounts through Plaid. It cannot move money, and Anthropic says ads and product pitches are ruled out.
- Cloudflare wants websites to know who is really behind an AI agent. Cloudflare says a store or bank sees only the assistant, never the person it works for, and that CAPTCHAs and narrow permissions do not fix it. It has started an industry group called Trusted Agent Delegation to work out what should.
Cheaper Every Quarter: Smaller Models, Smaller Bills
Three companies cut what it costs to get an answer, a clip or a code review out of an AI model.
- Thinking Machines’ second-generation Inkling is smaller and sharper. Inkling 2 Small runs 17B of its 594B parameters for each token and can read images. It ships on Tinker with closed weights first, and an Apache 2.0 release is promised for next week.
- Google’s video AI now makes 15-second clips for 7 cents a second. Gemini Omni 1.2 Flash raises the clip limit from 10 to 15 seconds and cuts the price 30%. Unlike the summer developer preview, it is now generally available.
- Devin’s AI code checker now costs about 25 to 75 cents a review. Cognition says Devin Review is 82 percent cheaper per pull request than in August, with the average catch up slightly to 1.32 bugs per review. The company gives no breakdown of what those bugs were or how many alerts were false.
Atoms and Balance Sheets: Robots, Simulators and the Money Behind Chips
Behind the models sit machines that move and the contracts that pay for the hardware.
- NVIDIA teaches robot vision models to plan a whole move at once. A small helper model drafts an entire path in one pass. NVIDIA’s researchers say robot planning runs about five times faster, with fewer misses on the target.
- NVIDIA lets anyone download four Cosmos simulators for cars and robots. The four 16-billion-parameter checkpoints on Hugging Face each do a different job, two for driving and two for robot arms, and the license allows commercial use. NVIDIA publishes no fidelity scores.
- How GPU rental contracts decide whether a data center gets financed. Ornn’s example provider cannot get a sensible loan until a hyperscaler signs for 640 chips over five years at $4 per GPU-hour. The signed contract, more than the hourly price, is what a lender prices.
Today’s Quick Hits
- Meta’s Muse app now talks back, with an avatar that reacts. Meta rolled out voice conversations and optional animated avatars in its Muse app on 5 November, starting in the US. Support for Meta’s glasses is still to come.
- Cursor now turns a described app into a live link. Cursor will build an app from a description, set up a database, test its security and deploy it to a live link. The company gave no pricing, limits or detail on how the checks work.
- Hugging Face publishes a guide to the AI design behind 18 of 20 top open models. Hugging Face’s handbook covers mixture-of-experts models, which wake only a few specialist blocks for each word fragment, and walks teams through training one on their own GPU cluster.