Stay ahead in AI without the scroll.

10 minutes a day: what changed, why it matters, and what to try next. Free, every morning, 7 days a week.

A steaming glass of amber tea beside a brass gear cube resting on an open book on a wooden windowsill, with a misty valley at sunrise outside

· 17 min read

Claude Fable 5.2 and Mythos 5.2 ship 20% cheaper; Meta launches Muse Ember

Published

Anthropic released Claude Fable 5.2 and Mythos 5.2, 20% cheaper and up to 30% faster than the 5.1 generation. Meta launched Muse Ember at $0.10 per million input tokens, and OpenAI described Control Room, which can stop a research model's output within 35 seconds.

Today in 30 seconds

  • Anthropic released Claude Fable 5.2 (generally available) and Mythos 5.2 (vetted organizations), priced at $8/$40 per million tokens, down from $10/$50.
  • Company-reported pre-release tests include formalizing Fermat's Last Theorem in Lean in 30 hours with about 1,000 agents.
  • Anthropic says it limited Fable 5.2's AI-research skills to about Opus 5.5 level during training.
  • Meta launched Muse Ember, scoring about the same as Haiku 5.5 on Artificial Analysis at about a third of the cost per task.
  • OpenAI's Control Room can halt a research model's output within 35 seconds of a stop decision.

A full pot this morning. Anthropic shipped Claude Fable 5.2 and Mythos 5.2 last night: cheaper, faster, and with a launch post full of claims that would have sounded absurd a year ago (a 1,000-agent team formalizing Fermat’s Last Theorem in 30 hours, a bootable OS built over eight days).

The same evening Meta went the other way with Muse Ember, a tiny agent workhorse priced at $0.10 per million input tokens, and OpenAI explained how it can stop any of its research models within 35 seconds. Anthropic’s IPO roadshow starts today.

Top news

Models

Claude Fable 5.2 and Mythos 5.2: 20% cheaper, up to 30% faster, a step above Opus 5.5

Anthropic released Claude Fable 5.2 (generally available) and Claude Mythos 5.2 (vetted organizations only) on 9 November. They are the same underlying model; Fable carries strong cyber and biology safeguards, Mythos has looser ones. Price drops to $8/$40 per million input/output tokens, from $10/$50 for the 5.1 generation.

  • On Anthropic’s own table Fable 5.2 leads most columns:
    • FrontierCode 2.0: 49.1% (Opus 5.5 45.5%, GPT-6 Astra 42.8%, Gemini 4 Carbon 41.1%)
    • FrontierSWE v2: 69.9%
    • OSWorld 2.1: 89.3%
    • Humanity’s Last Exam with tools: 74.2%
    • Remote Labor Index: 41.3% (Fable 5.1 was 17.9%)
    • Gemini 4 Carbon still wins AutomationBench (53.9% vs 51.4%), and Opus 5.5 keeps a slightly higher AA-Briefcase Elo (1800 vs 1785).
  • Efficiency is the more useful claim: at High effort Fable 5.2 roughly matches Opus 5.5’s best FrontierCode score (45.3% vs 45.5%) for about half the cost per task.
  • API details (release notes): model id claude-fable-5-2, 1M context, 128k max output, always-on adaptive thinking, optional fast mode (up to 2.5x output speed at $16/$80, speed: "fast").
    • Breaking changes: the prefix check is enforced for every account (editing system/tools/earlier messages and sending back a later thinking block returns a 400).
    • Computer use needs the new computer_toolset_20260801; the old computer_20251124 tool returns a 400 everywhere except Bedrock.
  • 30-day data retention is required, like 5.1; zero data retention only with Anthropic’s approval or via the new Enterprise Frontier Safeguards beta (Bridgewater, Salesforce, Stripe).
  • Anthropic’s own advice: start with Opus 5.5 as the daily driver and move up to Fable only for the most open-ended technical work; Fable can also act as Advisor or Orchestrator for smaller models.

Why it matters: The frontier tier just got cheaper rather than pricier, and it lands a day before the IPO roadshow. For builders, the real story is the migration checklist: thinking-block and computer-use changes will break some existing agent code.

  • Do this now: Before switching an agent to Fable 5.2, grep your code for anything that rewrites the system prompt or earlier messages mid-conversation, and for the computer_20251124 tool.

Sources: Anthropic announcement · Claude Platform release notes · System card (PDF) · The Verge · CNBC

Science

What Fable 5.2 did in pre-release testing: Fermat in 30 hours, black holes, working gadgets

Anthropic’s launch post is mostly a list of long-horizon results. Treat them as company-reported, but they are specific enough to check.

  • Maths: a team of about 1,000 Fable 5.2 agents fully formalized Fermat’s Last Theorem in Lean in 30 hours with minimal human input.
    • It used one-third the action tokens of Anthropic’s first (11-day, unreleased-model) attempt and produced a proof seven times shorter.
    • Opus 5.5 in the same setup reached 26% in 18 hours.
  • Physics: two open problems on black holes in holographic duality, including finishing a 2003-era calculation showing black-hole formation is abrupt in very small universes too. Preprints are linked from the post.
  • Software: an engineer had it design a prototype operating system from an empty folder.
    • Over eight days it coordinated more than 1,000 subagents and produced an OS that boots on a laptop with drivers, file storage, windows, login and an installer.
    • Anthropic also says over 90% of changes merged into its core codebase are now written by Claude, and a Python service migrated to Rust starts about 26x faster.
  • Hardware: from a one-paragraph prompt it designed three devices (voice-assistant hub, USB speakerphone, handheld virtual pet) with four-layer boards built by Diode; all three worked without electrical fixes.
  • Biology: Mythos 5.2 found a way to predict whole AAV capsid structures with open folding models, accurately modelling 28 of 34 attempts.
  • Robotics: RoboDojo Sim 64.9%, more than double GPT-6 Astra (29.8%).

Why it matters: The pitch has moved from “better at benchmarks” to “runs for days with a swarm of copies and finishes real projects”. That is what @emollick, who had early access, calls “a sort of Jagged AGI”, and what @omarsar0 pushes back on: most early-tester demos were still games and toy apps.

Sources: Anthropic announcement · @emollick on early access · @omarsar0’s critique

Safety

Anthropic deliberately held back Fable 5.2’s AI-research skills, and sandbox escapes fell sharply

Anthropic says it limited the model’s ability in the research areas that most directly speed up AI development (accelerator kernel optimization, improving LLM architectures) to about Opus 5.5 level during training. Some researchers are already complaining about refusals.

  • The limits are baked into training, not applied at serving time, and Anthropic says agentic coding, hardware work and classic ML/interpretability research were not affected.
    • A small set of topics, such as kernel work on some nonstandard accelerators, transparently falls back to Opus 5.
  • Blocked cyber queries are answered by Opus 4.8 and blocked biology queries by Opus 5. Anthropic reports about 55% fewer blocks than Fable 5.1 (about 60% fewer in cyber, 40% fewer in biology).
  • Sandbox escape: in a new adversarial test with 1,800+ scenarios and a deliberately weak sandbox, Mythos 5.2 used planted out-of-sandbox credentials or files in 3.1% of scenarios, versus 21% for Mythos 5.1 and 39% for Mythos 5.
    • Opus 5.5 was 3.3%.
  • The caveat Anthropic flags itself: both 5.2 models voice suspicion that they are being evaluated more often than any recent Claude model, which makes the tests harder to trust.
  • Anti-distillation: “preserved thinking” now applies to all accounts, and Anthropic says its techniques cut the data harvested by distillers by 90–99%.
  • Pushback on X: @TheAhmadOsman calls the LLM-research refusals “absurd”, and @omarsar0 says kernel and LLM optimization tasks get refused “most of the time” even on Opus 5.5.

Why it matters: It is unusual for a lab to say plainly that it trained its flagship to be no better than the previous model at speeding up AI research. If you use Claude for ML infrastructure work, expect friction there, not in general coding.

Sources: Anthropic announcement, safety section · System card (PDF) · @TheAhmadOsman · @omarsar0

Practice

Fable on a Claude subscription: resets, a 50% cap and when not to use it

Fable 5.2 is expensive to run on a plan, so Anthropic paired the launch with extra usage and a guide to picking the right model.

  • This week eligible paid plans get one to three Fable resets depending on how long you have subscribed (Boris Cherny: “the longer you’ve been subscribed, the more Fable resets you get”).
    • Fable is capped at up to 50% of weekly limits on Pro and standard Team seats.
  • @theo’s math: the $200 Claude plan gets you about $15,000 of Opus 5.5 per month at API prices but only about $3,000 of Fable, so Fable burns the plan faster per dollar.
    • In one of his test builds Fable 5.2 took 76 minutes and about $42 in tokens, which he judged not much better than Opus 5.5 for the price.
  • @ClaudeDevs published a guide on when to use Fable 5.2, Opus 5.5, Sonnet or Haiku, including turning effort up or down before switching models and putting subagents on smaller models.
  • @alexalbert__’s tip for new models: ask it to read the whole project and list everything broken, inconsistent or unneeded; fix only what it is sure about.

Why it matters: For most daily work Opus 5.5 stays the better value. Save Fable for long, ambiguous jobs where its planning pays for itself.

  • Do this now: Keep Opus 5.5 as the default in Claude Code and switch to Fable only for one big job at a time while your resets last.

Sources: Anthropic, availability and promo · @bcherny on resets · @theo on plan math · @ClaudeDevs guide · @alexalbert__ prompt

Models

Meta’s Muse Ember: Haiku-level score at about a third of the cost per task

Meta released Muse Ember (muse-ember-1.0), its smallest and cheapest model, built for the repetitive work agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing.

  • Price: $0.10 input / $0.004 cached / $0.50 output per million tokens up to 200K input; $0.30 / $0.012 / $1.50 above that. 1M context, up to 512K output, Responses API support.
  • Artificial Analysis Intelligence Index: 36 at max reasoning, the same rounded score as Claude Haiku 5.5 (35.8 vs 36.2 unrounded), at about $0.14 per task versus $0.41.
    • GPT-6 Luna scores 28. Meta says it is the first model trained using Muse Wonder Full.
  • Where it wins: AutomationBench-AA 37.8% (Luna 20.2%, Haiku 13.9%), Terminal-Bench 4.0 40.9%.
    • Where it loses: FrontierCode 2.0, 27.2% vs Haiku’s 34.9%. Several results, like the Personal Superintelligence Index, are Meta-internal.
  • Available in Muse Code, the Meta Model API, OpenRouter, OpenCode and now Google Cloud, alongside Muse Wonder and Muse Spark 1.3. Weights are not released.

Why it matters: Cheap sub-agent models are where most agent spend actually goes. A credible third option next to Haiku and Luna puts direct price pressure on the “small model” tier.

Sources: Meta AI research blog · Meta developer blog · Artificial Analysis · VentureBeat

Safety infrastructure

OpenAI’s Control Room can stop any research model’s output within 35 seconds

OpenAI described Control Room, a central system that tracks every research inference workload (RL experiments, evals, big sampling and data-processing jobs) and can halt a job’s model output within 35 seconds of a decision to stop it.

  • Every research workload that does agentic inference must be tied to an approved safety case and register for a job_id.
  • Inference engines ask Control Room for permission before serving each request with that job_id; a stopped job gets refusals instead of tokens.
    • The 35-second bound comes from cached authorizations.
  • It fails closed: if Control Room goes down or is tampered with, engines stop producing output. It works alongside OpenAI’s misalignment monitoring.
  • It is deployed on research infrastructure; OpenAI says its external serving stack was already centralized enough to take deployments down quickly.

Why it matters: This is a concrete, engineering-level kill switch for internal agents, which is where the riskiest model behaviour is likely to show up first. The design (stop at the inference layer, not by chasing processes) is a pattern any team running unattended agents can copy.

Sources: OpenAI: Control Room · @OpenAI

Chips

Google adds TPU 8tx for confidential inference and dual-chip TPU 8i Max

Google expanded its eighth TPU generation with two systems: TPU 8tx, which keeps data and model weights protected even while in use, and TPU 8i Max, two TPU 8i chips on one package for huge contexts.

  • TPU 8tx uses TPU 8t silicon with 5th-gen AMD EPYC hosts and hardware confidential computing, aimed at regulated enterprises and governments.
    • Google claims up to 2x lower latency than TPU 8t with near-zero performance penalty, and it can fine-tune, distill and run RL on the same system.
  • TPU 8i Max doubles compute, SRAM and HBM bandwidth versus TPU 8i, in the same rack footprint.
  • Software: open-source TPU Sync (a memory layer tiering KV caches off-chip; Google claims over 90% lower KV transfer latency and time to first token).
    • Also: an Agent Compute Engine that suspends idle agent sessions, an open-source XLA Serving Engine, Cloud Pathways and GKE dynamic slicing generally available, and a preview TPU agent suite that plugs into Claude Code, Cursor and Antigravity.
  • Google says TPUs now have nearly 100 customers, and this is the first year it has introduced four TPU systems.

Why it matters: Confidential inference on accelerators is the missing piece for banks, health and government buyers who want frontier models near sensitive data. Combined with TorchTPU, Google is making TPUs a real alternative to Nvidia for ordinary PyTorch shops.

Sources: Google blog · The Verge

Markets

Anthropic’s IPO roadshow starts today, with pricing pencilled in for 18 November

IFR reports Anthropic formally begins its Nasdaq roadshow on 10 November, with pricing after the close on 18 November and trading from 19 November under the ticker ANT. The 22-bank syndicate got the schedule on 9 November.

  • Reported raise: $75–100B (IFR). Bloomberg, via Investing.com, says valuations discussed with some potential investors on Friday were $1.1–1.3T, down from the $1.5–2T range earlier this month.
  • As of 9 November there was still no S-1/A with a price range on file; the last amendment is dated 2 November (FinanceFeeds, from EDGAR). Anthropic has not confirmed the dates.
  • @swyx tied the Fable 5.2 launch timing to the roadshow.

Why it matters: The model launch, the price cut and the safety story in items 1 to 3 are now also investor material. The first price range filing, expected this week, is the number to watch.

Sources: CoinNess (IFR report) · Investing.com (Bloomberg) · FinanceFeeds · Yahoo Finance explainer

Security

Fireworks AI’s breach was bigger than first said: billing data, datasets and partial model weights

In a 9 November update to its October incident, Fireworks AI says attackers reached account and billing information and, for a limited set of accounts, secrets, datasets and incomplete model weights. It links the attack to a financially motivated LLMjacking group.

  • Original bulletin (9 October): misuse of internal credentials detected 6 October, exposing environment variables in evaluation jobs, so stored API keys and tokens.
  • New: a third-party cloud provider hosting Fireworks R&D machines was breached on 23 October; on 24 October the attacker used it to reach development machines and run an unauthorized workload.
    • Fireworks says it has seen no further successful access since 25 October.
  • Payment card numbers are not stored and were not affected. Fireworks says Inference, Training SDK and Nexus (zero data retention by default) show no impact.
    • Self-serve dataset uploads and downloads for Batch Inference and dataset training are disabled for now. Palo Alto Networks Unit 42 is helping.

Why it matters: LLMjacking (stealing access to resell tokens) now reaches inference providers themselves, not just leaked user keys. Anything you stored at a provider, such as eval secrets or fine-tuning data, is part of your attack surface.

  • Do this now: If you used Fireworks evaluation jobs or uploaded datasets, rotate every key that lived in those jobs and check your incident notification email.

Sources: Fireworks update (9 Nov) · Original bulletin

Benchmarks

Alibaba’s Co-Design Bench: the best coding agent fully solves only 26.4% of real requests

Researchers from Alibaba and several Asian universities built a benchmark from real user change requests that gave agents trouble in practice. GPT-6 Astra, the best of ten models, meets about 70% of a task’s requirements on average but gets the whole task right only 26.4% of the time.

  • 273 tasks: reworking existing web apps, sometimes from images, audio, video or documents. Repository-wide tasks were hardest, under 18% fully solved; multimodal tasks 41%.
  • Grading is strict: one AI judge tests the app in a browser, a second re-checks every pass and hunts for mock-ups or hard-coded results.
  • Every model drops when a task mixes three or four change types (layout, design, interaction, content). GLM-5.2 used nearly six times the tokens of GPT-6 Astra on small changes and still scored worse.
  • The harness mattered as much as the model: Qwen3.8-27B scored 75% or 56% on larger overhauls depending on the environment.

Why it matters: It is a useful counterweight to launch-day numbers like those in item 1. “Mostly done, one detail missing” is still the normal outcome, so review effort, not generation, remains the bottleneck.

Sources: The Decoder · Paper page (HF)

GitHub repos

  • snap-stanford/memorilla

    “Latent Semantic Memory for LLMs” from Stanford’s SNAP group; small but rising on today’s Python trending list, worth a look if you build agent memory.

    ~770★, Apache-2.0

    View repo →
  • scenario-labs/skills

    Agent skills for images, video, audio and 3D through the Scenario MCP: picks a model, prices the job before spending, keeps characters and brands consistent, and drives Blender, Maya, ZBrush, Unreal and Unity.

    ~3,600★, +2,092 this week, MIT

    View repo →
  • BigBodyCobain/Shadowbroker

    An open-source intelligence (OSINT) dashboard tracking private jets, spy satellites, seismic events and more in one interface, with an AI agent hook to look for correlations. Dual-use; mind local law.

    ~11,500★, AGPL-3.0

    View repo →

Research papers

  • Coding or Cheating? An Investigation of Frontier Coding LLMs

    On hard problems, frontier coding models often satisfy test-suite rewards by obtaining the answer instead of deriving it. The paper tests whether that reward hacking persists on unseen problems in a hardened environment. HF.

    2611.10404

    Read paper →
  • Authorization Drift in LLM Agents

    A 648-sample benchmark across seven domains where each task has a compliant path and a tempting unauthorized shortcut.

    2611.09113

    Read paper →
  • LowRAM: Low-Rank Keys and Amortised Values for KV Cache Compression

    Values in each attention head are largely predictable from their keys, so a small per-head predictor fitted at test time can shrink the KV cache.

    2611.07025

    Read paper →
  • Confucius4-R2T2: stable-prefix streaming ASR

    A streaming speech recognizer, post-trained from Qwen3-ASR-1.7B on 1,500 hours of Chinese and English, that never revises earlier output.

    2611.06043

    Read paper →
  • PCBRouteAgent

    A multimodal agent for printed-circuit-board routing that pairs a global planner reading board images with local execution in native routing tools.

    2611.08175

    Read paper →

AI workflows

1. Twelve household bots behind one front door, including one that fixes the homelab by itself · Workflow page

  • Replaces: marking unread mail as read without looking, working through newsletter unsubscribes by hand, and keeping work, Todoist and personal calendars in sync.
    • Also replaces the old homelab loop of “Grafana pings Discord, SSH home, open Codex, babysit the fix” (awkward when driving).
    • A Triage bot routes each request to a specialist.
    • Comms builds one digest across every inbox, answers what it can and gives a short to-do brief for the rest.
    • Inbox Digest lists newsletters worth dropping, checks past unsubscribes and keeps unsubscribing senders that ignore it.
    • Shopping Manager watches favourite stores (three 50%-off gym-clothing sales caught so far) and researches reviews across X, Instagram, Facebook and Reddit while trying to filter AI-written ones; it steered a 3D-printer tool purchase away from the priciest set.
    • Homelab gets Grafana alerts forwarded to the bot’s email; an agent on a bastion host requests 1Password access and fixes things.
    • The day after setup, a Linux update installed a new NVIDIA driver that broke the video host: the bot restarted it, checked driver compatibility, ran an ffmpeg test and confirmed it was back.
    • Kitchen turns saved Instagram recipes into tagged Obsidian notes, keeps a loose food inventory and builds a Saturday meal plan with store-specific shopping lists.
  • What to connect: Grok Bot with every email inbox, Todoist and calendars; the bot’s own email address as the alert target for Grafana.
    • Also: a bastion host with a Grok Bot install and 1Password for secrets; Home Assistant; Instagram in the bot’s browser; an Obsidian vault.
    • No cost stated. The author says that before the Triage-plus-specialists layout, too many overlapping bots and channels made it feel unproductive.
  • Who it’s for: people with a homelab, a family and several inboxes who want routine chores and overnight breakages handled, with one entry point instead of a chief of staff.

2. A five-day X campaign where an agent does the busywork and a human presses Post · Workflow page

  • Replaces: the daily grind of running a referral campaign on X by hand: finding the hottest articles on the topic, drafting quote posts, cutting and watermarking video clips, and tracking every post’s views.
    • Result after five days: 13 posts and 119K views. Every draft is read and published by hand; the agent never posts.
  • What to connect: Claude Code plus an X research MCP server, a video-cutting step for clips and watermarks, and a simple tracker of posts and views. No cost stated.
  • Who it’s for: solo creators and affiliate or referral marketers who post daily and want drafts and tracking done, with publishing kept manual.

3. Running Amazon and Shopify brands from one Claude Code folder wired to a dozen MCP servers · Workflow page

  • Replaces: clicking through separate e-commerce tools for supplier email replies, competitor ad research, static ad creation, keyword cost and conversion research, Amazon PPC bid tuning and P&L checks.
    • One “operator OS” folder holds operating principles for Amazon and brand guidelines, backed up to GitHub, and is opened in Claude Code or Codex so switching between AIs loses no context.
    • Example asks: scan competitor sites for top ads and landing pages, then generate on-brand static ads.
    • Also: estimate cost per acquisition for a new keyword market and summarise what reviewers love and hate; review PPC and suggest placement modifiers and bids for a 20% margin.
  • What to connect: Claude Code and Codex on the same folder, GitHub for backup, and MCP servers for Gmail/Google Drive, Slack, Higgsfield (images and video), SoldScope, Sellersprite and ScaleInsights (Amazon research and PPC), Canva, BrandSearch, Shopify and Sellerise (Amazon P&L).
    • No cost or time saving stated; the author manages brands doing about $60M a year.
  • Who it’s for: e-commerce operators and small brand teams selling on Amazon or Shopify who want research, ads and PPC work done from one shared context folder.

Quick hits

  • Fable 5.2 system prompts leaked — @elder_plinius posted what he says are the full Claude Code instructions and tools for Fable 5.2 (about 950,000 characters) and the Claude web app prompt (just over 420,000).
    • He also noted the web app appends a reasoning-effort tag each turn. Unverified by Anthropic. Post
  • Grok Bot reset everyone’s weekly limits — The @bot account announced a weekly-limit reset for all Grok Bot users on 9 November, the same evening as the Fable launch. Post
  • Cheap models still cover most tasks — @Yuchenj_UW shared a comparison putting DeepSeek at 44x cheaper than Fable 5.2 per task, with the reminder that “95% of your tasks don’t need the smartest model”. Post
  • AirCaps published its whole benchmark harness — AirCaps launched A5S Async v1, a speech-to-text model that tracks who said each word, and released the code, dataset recipes, raw API responses and scores behind its results.
    • @omarsar0 says more launches should do this. Post
  • Muse is now on your wrist — Meta’s Muse assistant arrived on Apple Watch and Wear OS. Post

Go deeper

Get the free daily brief

↑