Anthropic just shipped its new top model. Claude Fable 5.2 is out today, 20% cheaper than Fable 5.1 and up to 30% faster, and it tops Anthropic's own scoreboard against OpenAI's GPT-6 Astra on coding, research and real-world work.

It also lands weeks before Anthropic's planned IPO, which makes the price cut as interesting as the benchmarks.

Also tonight: Meta gives every Muse and Meta AI user a smarter model, Nous Research ships an open model with an agenda, and four tools worth trying this week.

Claude Fable 5.2 is live today in Claude (included for up to 50% of weekly limits on Pro and standard Team seats), the API (claude-fable-5-2), AWS, Google Cloud and Microsoft Foundry. It costs $8 per million input tokens and $40 per million output, down 20% from Fable 5.1, and Anthropic says answers come back up to 30% faster. It handles text and images, with a 1M-token context window and up to 128K tokens of output.

On Anthropic's own tests it beats both Claude Opus 5.5 and OpenAI's GPT-6 Astra across agentic coding, research and real-world work, with the biggest jump on the Remote Labor Index, which roughly doubled from Fable 5.1. The showpiece: about 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in roughly 30 hours. Its twin, Claude Mythos 5.2, is the same model with looser cyber and biology safeguards, open only to vetted security and life-sciences teams.

Our take: this is the strongest model you can actually use today, and the price cut makes it easier to justify for long, hard jobs like multi-hour agent runs or research that ends in a finished document. But it isn't your new default. Fable thinks hard on every request (adaptive thinking can't be turned off), and Anthropic itself says to start with Opus 5.5 and step up only when a task needs it. Fable also refuses about 55% fewer harmless requests than 5.1, which should cut the frustration.

Try it: Claude app · API docs and pricing

On paper, Fable 5.2 undercuts OpenAI. It lists at $8 per million input tokens and $40 per million output tokens, against $10 and $50 for GPT-6 Astra, and cached reads cost $0.20 per million versus Astra's $1.00, Fortune notes. For long documents and repeated context, that cache discount is the bigger deal.

But per-task cost tells a different story. Artificial Analysis ranks Fable 5.2 first on its Intelligence Index, yet at its top setting it cost about $26.50 per task versus about $10 for Claude Opus 5.5, per The Decoder. That's why Anthropic itself still says to start with Opus 5.5 for most work and step up to Fable for the hard stuff. The timing matters too: Anthropic wants to list before Thanksgiving, and a flagship that beats GPT-6 Astra at a lower sticker price is exactly the story it wants investors reading.

🧠 Meta Brings Muse Wonder Smarts To Every User

Meta's top model, Muse Wonder, has mostly been for paying users and developers since October. Now Meta is rolling out specialized versions of its existing models, trained using Muse Wonder, to everyone on the Muse agent app and Meta AI. It arrives automatically.

The focus is memory. Meta says the new Muse model correctly uses critical context in 90% of tasks, up from 76%, forgets goal-critical details 2.4x less often, and does 67% better on everyday tasks where something goes wrong. All of these are Meta's own tests.

Our take: memory is where consumer assistants win or lose, and giving a frontier-trained model to free users is Meta's edge over rivals that charge for their best tier. Better memory also means more personal data kept, so ask Meta AI "What do you remember about me?" and prune what you don't want stored.

Try it: Meta AI · Muse Spaces

Nous Research released Arete-27B on Hugging Face over the weekend under Apache 2.0. It's a fine-tune of Alibaba's Qwen3.8-27B, trained with supervised fine-tuning only on about 52.8 million tokens. The goal is unusual and stated openly: shift how a Chinese open model handles politically sensitive topics toward Western values.

On Nous' own 803-prompt Propaganda_Eval, answers tagged as propaganda fell from 22% to 0.9%, while its general capability scores barely moved. These are self-reported numbers on an eval Nous built and released alongside the model.

Our take: many teams avoid strong, cheap Chinese open models because of how they answer on certain topics. Arete tests whether a small post-training pass can fix that without breaking the model. "Western values" is a choice, not a neutral fix, so run your own evals before shipping it.

Try it: Weights on Hugging Face · Propaganda_Eval dataset

The other stories worth your attention:

Alibaba tested coding agents on real app changes, and even the best of ten models, GPT-6 Astra, fully got only a small share of tasks right. See how big the gap is with older benchmarks. (The Decoder)

A startup says AI cracked a 2008 quantum-computing conjecture using a non-public Anthropic model, with the proof checked in Lean. Stanford researchers are still reviewing the draft. (PSI)

Nvidia's board opened an independent investigation into Jensen Huang after reports he was targeted by a Chinese spy operation. Here's what Huang admitted in his memo. (TechCrunch)

Google's smartwatch AI spots hidden high blood pressure without a cuff, and in a Nature study it beat a standard office reading. See how it works. (Google Research)

OpenAI reportedly banked $11 billion at a $1.25 trillion valuation, but it's less than it set out to raise. Here's the gap. (TechCrunch)

Anthropic's Claude Fable 5.2 is out today, 20% cheaper than Fable 5.1, up to 30% faster, and ahead of GPT-6 Astra and Opus 5.5 on Anthropic's own coding and research benchmarks. The catch is cost per task: it can still burn far more than Opus 5.5, which Anthropic recommends for everyday work. Meta pushed Muse Wonder-trained models to every Muse and Meta AI user, and Nous Research's Arete-27B shows how a cheap fine-tune can reshape an open model's views. The thread: the frontier is getting cheaper per token, so the winners will be the people who pick the right model for each job.

Stay building. 🤖

Our newest video is up. Watch it below.