Claude Fable 5.2 ships 20% cheaper per token and tops Artificial Analysis

Today leads with Anthropic's new top model because it's what developers spent the night arguing about. Two Hacker News threads on it passed 700 and 900 points, and the loudest question was whether it's really cheaper. We checked an outside test. Also today: Claude Pro subscribers get the new model for a week. Anthropic's own guide has a one-line prompt that gets it to tidy code. Meta shipped a small, cheap model for agents. Anthropic put a price range on its IPO. And Google DeepMind's agents proved the code behind Signal's encryption secure.

What changed this week

Claude Fable 5.2 ships 20% cheaper per token and tops Artificial Analysis

Anthropic released Claude Fable 5.2, its model tier above Opus, on Nov 9 at $8 per million input tokens and $40 per million output, down from Fable 5.1's $10 and $50. It's now Claude Code's default Fable model, and Anthropic says it answers "up to 30% faster." Mythos 5.2 is the same model with looser cyber and biology safeguards, for verified organizations only. On FrontierCode 2.0, which grades whether open-source maintainers would merge a model's changes, Anthropic reports 49.1% at its Xhigh effort setting, and at High about level with Opus 5.5's best "for roughly half the cost per task." Artificial Analysis, which re-runs models itself, ranks Fable 5.2 first of 74 as of 12:19 UTC Nov 10. At max effort it wrote 275 million output tokens to 5.1's 163 million, so each task cost $26.54 against $12.99. Cheaper tokens can still mean a bigger bill. Try High before Max.

Claude Pro plans get Fable 5.2 through Nov 16, plus one limit reset

Anthropic's launch offer puts every Fable model inside Pro plans and standard Team seats until 11:59 pm Pacific on Nov 16, free up to half of each weekly limit. Those accounts also get one Fable limit reset, which returns the weekly usage spent on Fable. Max and Ultra plans get one to three resets by subscription length, good until March 8, 2027. Resets need paid billing before Nov 9. Pick Fable in the model picker.

Claude Fable 5.2 launch offers | Claude Help Center

Claude Help Center

Anthropic says telling Fable 5.2 "this codebase is yours" gets it to clean up code

Anthropic's prompting guide for Fable 5.2 says the model listens to who the prompt says owns the code. Told nothing, it tends to add to the existing structure, "even when that repeated existing code." Given the line "This codebase is yours," it merged repeated code "in about half of the tries" on small coding tasks, against "under a tenth" without it. The cost is about 1.5 times the output tokens and changes nobody asked for. Use the line on your own repos. Then "review the diff and not only the summary."

Prompting Claude Fable 5.2

Prompting patterns specific to Claude Fable 5.2 and Claude Mythos 5.2: effort, early stopping, approvals, code ownership, progress updates, subagents, caution that builds up, instructions written for earlier models, final reports, and refusals.

Claude Platform Docs

Meta's Muse Ember undercuts Claude Haiku 5.5 on price, at $0.10 per million input tokens

Meta released Muse Ember on Nov 9, the smallest of its Muse models, built for "tool calls, coding subtasks, routing" and other work agents repeat all day. It costs $0.10 per million input tokens and $0.50 per million output, in Meta's API and Muse Code. Meta says it "matches Haiku's best score at about a third of the cost" on Artificial Analysis's overall index, against Anthropic's smallest model, Haiku 5.5. Mergeable code is weaker. FrontierCode 2.0 has Haiku at 34.9, Ember 27.2. Try it for subagents.

Introducing Our Smallest, Most Efficient Model: Muse Ember

Muse Ember is our smallest and most efficient model, designed for complex reasoning execution and the work agents repeat all day.

Meta AIMeta AI Research

Anthropic sets its IPO at $70 to $78 a share, to list on Nasdaq as ANT

Anthropic opened its IPO roadshow on Nov 10, the pitch to investors before shares trade. It's offering 825,000,000 Class A shares at an expected $70 to $78 each, on Nasdaq under the symbol "ANT." Anthropic sells 801,351,354 of them and existing stockholders sell the rest. The announcement gives no trading date. It names no change to Claude's products or prices.

Anthropic launches roadshow for proposed initial public offering

Today Anthropic, PBC launched the roadshow for its proposed initial public offering of its Class A common stock.

AnthropicAnthropic

Google DeepMind's agents prove Signal's encryption code secure, in four weeks

Google DeepMind says AI agents proved key security properties of the actual code behind the Signal protocol, which protects chats in Signal and Google Messages. Researchers wrote the specifications; agents running Google's Gemini Carbon model generated "over six million lines of Lean," a proof language whose checker confirms each step, and flagged "just one percent" for humans. It took four weeks. When the team put two previously fixed bugs back, the system proved each buggy version insecure. The post names no product to install.

Formal Verification of Signal’s Protocol with Gemini

AlphaProof Code agents wrote over three million lines of machine-checked proof in Lean to verify the core cryptographic protections in Signal down to machine instructions.

Google DeepMindPushmeet Kohli and Swarat Chaudhuri

Also worth your time

• Ethan Mollick, who had early access, calls Fable 5.2 "Jagged AGI" and shows where it still falls short
• A developer had Fable 5.2 build his nephew's beat 'em up game from the boy's drawings
• Fable 5.2 is now in GitHub Copilot for Pro+, Max, Business and Enterprise users, billed at list price
• Claude Code v2.1.321 fixes MCP sign-ins that sent refresh tokens to another host after a redirect
• Fireworks AI says attackers reached some customers' datasets and model weights; Notion says AI chat transcripts were exposed
• JPMorganChase extends Claude Code past its engineers, per Anthropic
• Codex CLI 0.167.0, out Nov 6, adds an agent command center, Google Cloud Vertex AI setup and switching between connected backends

Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.

The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.