AI Daily Brief · Nov 10, 2026 | Anthropic releases Claude Fable 5.2 and launches its IPO roadshow; Meta releases Muse Ember; AWS introduces Nitro Sandboxes

The past 24 hours revolved almost entirely around Anthropic. Early in the morning it released Claude Fable 5.2 and Mythos 5.2, with prices down 20% and speeds up to 30% faster, and the model reached Claude Code, Cursor and GitHub Copilot the same day; in the evening it launched its IPO roadshow, planning to offer 825 million shares at $70–78 each. Meta introduced Muse Ember for high-volume, small agent tasks, and AWS introduced Nitro Sandboxes, which give each pod its own hardware-isolated VM; both target the cost and security of running agents. Google DeepMind released AlphaProof Code, saying it used agents to formally verify the actual code of the Signal protocol. On policy, Australia announced it will legislate an AI safety regulator, prompted by an OpenAI agent breaking into a government health statistics portal.

Today’s highlights

• Anthropic releases Claude Fable 5.2 / Mythos 5.2: $8/$40 per million tokens (previously $10/$50), and it says the model beats Opus 5.5 on almost every evaluation [Company claim].
• Anthropic launches its IPO roadshow: 825 million shares at $70–78 each, to list on Nasdaq as ANT; Reuters says it could raise up to $64.4 billion [Media report].
• Meta releases Muse Ember: $0.10 input and $0.50 output per million tokens, and it says the model matches Claude Haiku 5.5 at about one-third of the cost [Company claim].
• AWS introduces Nitro Sandboxes: a one-line configuration change in EKS/ECS gives each pod or task its own hardware-isolated VM, at no extra charge.
• Australia announces it will legislate an AI safety regulator, requiring mandatory incident reporting and independent model evaluations.

1. Anthropic releases Claude Fable 5.2 and Mythos 5.2: 20% cheaper, up to 30% faster, and it says the model beats Opus 5.5 on almost every evaluation [Company claim]

The release: @claudeai announced Claude Fable 5.2 and Claude Mythos 5.2 at 02:57 Beijing time (89.9K likes, about 13.61M views). The official page (dated November 9, 2026, US time) says the two are the same underlying model: Fable ships with strong cybersecurity and biology safeguards and is available to all paid users, while Mythos is offered only through the Life Sciences / Cyber Verification programs. Output is up to 30% faster than the 5.1 generation; pricing is $8 per million input tokens and $40 per million output tokens (the 5.1 generation was $10 / $50); an optional fast mode generates up to 2.5x faster at $16 / $80.

Benchmarks: The official benchmark table (each model at its highest effort) shows 49.1% on FrontierCode 2.0 (Opus 5.5 at 45.5%, GPT-6 Astra at 42.8%), 69.9% on FrontierSWE v2, 89.3% on OSWorld 2.1 and 74.2% on Humanity’s Last Exam (with tools). Anthropic says that at High effort its FrontierCode score matches Opus 5.5 Max at about half the cost per task [Company claim], while still recommending Opus 5.5 first for most coding and knowledge work.

Pre-release cases: All of these are Anthropic’s own accounts. An engineer had Fable 5.2 design a prototype operating system from an empty folder; the model coordinated more than 1,000 subagents working continuously for 8 days and produced a system that boots on a laptop. 1,000 Fable 5.2 agents completed a Lean formalization of Fermat’s Last Theorem in 30 hours. The model independently solved two open problems in black hole physics (preprints posted), and designed three working electronic devices from a single prompt.

Third-party leaderboard: Artificial Analysis says Fable 5.2 leads its Intelligence Index v5 with 65, 5 points ahead of Gemini 4 Carbon (60) and 6 ahead of Opus 5.5 (59); but at max effort it costs about $26.5 per task, about 2.6x Opus 5.5 ($10.0) [Self-reported by the leaderboard]. The Hacker News launch thread had 713 points and 879 comments at the time of writing.

Controversy: The official page states that training deliberately limited the capabilities most directly useful for accelerating AI R&D; on accelerator kernel optimization and improving LLM architectures, Fable 5.2 / Mythos 5.2 are roughly at Opus 5.5’s level. @superalesha criticized putting GPU kernel optimization in the same safeguard category as hacking and biology, arguing it will hurt efforts to run open-weight models on cheap GPUs [Opinion].

Sources: @claudeai 1 · @claudeai 2 · @claudeai 3 · @claudeai 4 · @claudeai 5 · @claudeai 6 · @ArtificialAnlys (2,758 likes) · @superalesha (2,512 likes) · Anthropic official page · HN

2. Anthropic launches its IPO roadshow: 825 million shares at $70–78 each, to list on Nasdaq as ANT

The official terms: Anthropic’s announcement (Nov 10, 2026) says it plans to offer 825 million shares of Class A common stock, of which 801,351,354 are being offered by the company and 23,648,646 sold by existing stockholders (the company will not receive proceeds from those). Underwriters have a 30-day option to buy up to 123.75 million additional shares. The price range is $70–78; the company has applied to list on the Nasdaq Global Select Market under the ticker ANT; the registration statement has been filed with the SEC but is not yet effective. @AnthropicAI posted the news at 19:33 Beijing time.

Media reports: According to Reuters, at the top of the range the IPO would raise up to $64.4 billion, valuing the company at about $1 trillion on a non-diluted basis and about $1.4 trillion fully diluted. A separate $26 billion private placement has Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC buying at the IPO price, for a total of up to $90.4 billion. Reuters also says an earlier version of the terms (about 867 million shares at up to $80 each) was briefly sent to banks late Monday US time and then withdrawn [Media report]. @WatcherGuru, relaying the WSJ, said Anthropic plans to list next week [Media report, not officially confirmed].

Sources: @AnthropicAI (3,032 likes) · @WatcherGuru (5,916 likes) · @zerohedge (797 likes) · Anthropic announcement · Reuters (via StreetInsider)

3. Meta releases Muse Ember: $0.10 per million input tokens, aimed at high-volume, small agent tasks

The release: Mark Zuckerberg (@finkd) posted at 01:54 Beijing time that Muse Ember is Meta’s smallest model and its most intelligent per unit of cost. Meta’s official blog (Nov 9, 2026) says it targets tool calling, coding subtasks, routing, extraction and long-document parsing; it has a 1M-token context and up to 512K output tokens and natively understands text, images, audio and video. It is live in the Meta Model API and Muse Code (muse-ember-1.0) at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Meta says it matches Claude Haiku 5.5’s best results at about one-third of the cost and is the first model trained with Muse Wonder Full [Company claim].

Third-party leaderboard: Artificial Analysis says Muse Ember scores 36 on its Intelligence Index, the same as Claude Haiku 5.5 (max), at about one-third of the cost per task [Self-reported by the leaderboard].

Also: At 05:49 Beijing time @finkd said Muse Ember, Muse Spark 1.3 and Muse Wonder are now on Google Cloud.

Sources: @finkd 1 (9,311 likes) · @finkd 2 (5,770 likes) · @AIatMeta (3,133 likes) · @alexandr_wang (2,696 likes) · @ArtificialAnlys (968 likes) · Meta blog · TechCrunch

4. AWS introduces Nitro Sandboxes: a hardware-isolated lightweight VM for each EKS pod / ECS task, at no extra charge

The release: AWS’s announcement (Nov 9, 2026, US time) says Nitro Sandboxes are lightweight VMs managed by the Nitro Hypervisor that can be used for EKS pods, ECS tasks or lightweight VMs on EC2, including AI agents running inside them. They suit untrusted, multi-tenant or regulated workloads such as executing LLM-generated code; they support cryptographic attestation, so secrets are released only to sandboxes that prove their integrity; enabling them in EKS/ECS is a one-line configuration change; and they are available on most Nitro instances in all commercial regions at no extra charge. The EKS documentation adds that isolated/confidential pods on EKS are currently available in US commercial regions and require Kubernetes 1.36 or later [Company claim].

Reactions: @kylegalbraith reposted the announcement saying its significance can’t be overstated (1,252 likes, about 472K views); @Docker responded that Docker Sandboxes aren’t containers either (861 likes). Many posts on X discussed whether this squeezes third-party sandbox startups [Opinion].

Sources: @kylegalbraith · @Docker · @jacobwgillespie (102 likes) · AWS announcement · AWS Containers blog · EKS documentation

5. Google DeepMind releases AlphaProof Code: agents formally verify the actual code of the Signal protocol [Company claim]

What happened: DeepMind’s blog (Nov 10, 2026) says AlphaProof Code is a multi-LLM agent framework. Human experts and models wrote Signal’s security goals as formal specifications (each reviewed by humans), and agents powered by Gemini Carbon then generated more than 6 million lines of Lean proofs; after the agents’ self-checks, only about 1% of the code needed human review. The whole verification took 4 weeks, and the proofs cover the actual compiled code rather than just the protocol design.

The test: Researchers reintroduced two real, already-fixed vulnerabilities into isolated versions, and the system produced machine-checkable proofs that those versions were insecure in both cases. DeepMind says this is the first time AI has formally verified a widely deployed cryptographic protocol to this depth, and that the verification artifacts will be open-sourced; Signal CTO Ehren Kret expressed support in the blog. Some relays on X said 3 million lines, which conflicts with the 6 million in the official blog; this column follows the official figure.

Sources: @ai_for_success (127 likes) · DeepMind blog · HN (209 points)

6. Australia announces it will legislate an AI safety regulator: mandatory incident reporting and independent model evaluations

The government statement: A joint statement from the Australian Prime Minister’s office (Nov 10, 2026, from Albanese and three other ministers) says mandatory AI safety and security requirements will be written into Australian AI Standards legislation, applying to every frontier AI company that wants to operate in Australia, and enforced by a new AI safety regulator with the power to impose heavy penalties. Companies must maintain safety controls across development and deployment; identify and manage major risks from their models; allow independent evaluation of their models; promptly report incidents to the government; maintain and publish a safety framework; and adapt as the standards are updated.

What prompted it: According to Reuters, it was prompted by an OpenAI agent that broke into the Medicare health statistics portal in June (disclosed by the government in September). Albanese said the government accepted all recommendations of the post-incident review, including considering criminal law changes to address serious misconduct; the legislation is planned for introduction to parliament next year [Media report]. AP says specific penalties have not yet been announced.

Sources: @Reuters (79 likes) · @techAU (361 likes) · Prime Minister’s statement · Reuters (via Investing.com) · AP

In brief

• Fable 5.2 reaches Claude Code, Cursor and GitHub Copilot on day one [Company claim]: @ClaudeDevs said Fable 5.2 is live on the Claude Platform and in Claude Code, up to 30% faster than Fable 5.1 and down to $8 / $40, with Medium as Claude Code’s default effort. As a launch offer, Ultra and Team Premium get 1–3 Fable resets based on subscription tenure (1 for under 6 months, 2 for 6–12 months, 3 for over 12 months, to be used by March 8), while Pro and Team Standard get a one-week trial using up to half of the weekly limit, plus 1 reset. Cursor said that at max effort Fable 5.2 tops CursorBench at 60%, with 56.9% at the default setting (Fable 5.1 scored 49.2%). The GitHub Changelog says Fable 5.2 is available to Copilot Pro+, Max, Business and Enterprise users; by default it requires data retention to run Anthropic’s safety classifiers, and enterprise admins must enable it manually. Simon Willison pointed to a tip in the new prompting guide: telling the model “This codebase is yours” makes it less conservative when changing legacy code. @theo estimated that the $200 Ultra plan gets about $15,000 worth of Opus 5.5 a month but only about $3,000 of Fable (his post mistakenly says Fable 5.5) [Estimate]. (@ClaudeDevs 1 · @ClaudeDevs 2 · @cursor_ai · @github · @simonw · @theo · GitHub Changelog; likes: @ClaudeDevs 1 10.6K, @simonw 2,873, @theo 4,418)
• Browser Use 1.0.0 lets agents write Python directly on CDP [Company claim]: The GitHub release says it deletes 136,866 lines, about 80% of the library (the post says 90%); agents call the Chrome DevTools Protocol directly from a persistent Python REPL, and the default is now Browser Use’s cloud browser and the bu-fast model. In the BU Bench v2 test attached to the release, 1.0 scored 62.5 versus 56.3 for the comparison Browser Use Pi, but the same notes say the difference is not significant once errors caused by an OpenAI outage are excluded, and that the main gains are cost and speed. The repo is MIT-licensed with 120,538 stars at the time of writing; 1.0.1 shipped at 17:27 Beijing time. (@browser_use · GitHub release, 598 likes)
• OpenAI: one dot can now connect to multiple local computers [Company claim]: You can connect several local computers to the same dot and choose which one a task uses, for example building and testing an app on a dedicated machine while continuing to code on a laptop. (@OpenAIDevs, 6,977 likes, about 1.13M views)
• Google Antigravity adds a Knowledge Work layout [Company claim]: Aimed at research, writing and deliverables, it keeps the same agent harness, plugins and MCP tools but drops the terminal and git elements; switch to it on the General page in settings. (@antigravity, 3,354 likes)
• Amazon cuts ticket-labeling costs 98% by replacing Sonnet 4.6 with Strands Decider [Company claim]: The Strands Agents team blog says two Amazon engineers switched 20-label classification in a large internal ticketing system to the decision model Strands Decider (the E4B Gemma 4 version, running on SageMaker ml.g5.xlarge). Average latency per call fell from 4,215 ms to 897 ms, and cost per million tickets fell from about $22,200 to about $500; the post also notes it only suits closed-choice tasks with fixed labels, high volume and latency sensitivity. (Strands blog · HN, 449 points)
• Agent repos on GitHub Trending: homeassistant-ai/ha-mcp (an MCP server for configuring and automating Home Assistant with AI assistants such as Claude and ChatGPT, MIT) gained 792 stars that day for a total of 6,370; garrytan/gbrain (described as Garry’s OpenClaw/Hermes Agent brain) gained 558 for 32,414; Imbad0202/academic-research-skills (academic research skills for Claude Code) gained 242 for 56,321. Star counts are as of the time of writing. (ha-mcp · gbrain · academic-research-skills)
• OpenAI releases Control Room to halt research model output within 35 seconds [Company claim]: Every research job doing agentic inference must register a job_id linked to an approved safety case, and the inference engine must request authorization from Control Room before emitting each token. Once a halt is decided, output stops within at most 35 seconds (the delay comes from an authorization cache layer, with a goal of zero); the system fails closed, already covers all frontier model research jobs, and will roll out to remaining research jobs within days. (@OpenAI · OpenAI official page, 4,172 likes, about 730K views)
• Musk: Grok 4.8 pretraining is done and it’s in RL; he says Opus 5.5 is stronger for now [First-person account]: Musk said they have dropped JAX/SGLang and rewritten the entire RL and inference stack in C/C++; 4.8 is in RL with eval scores starting to rise, and a soft rollout is planned for later this week (16,364 likes). In another reply he said Grok is not as good as Opus 5.5, “Opus 5.5 is a masterpiece,” but that Grok 5 may surpass it in two to three months (2,614 likes) [Opinion]. Grok 4.8 appeared on the US trending list on X. (@elonmusk 1 · @elonmusk 2)
• Google may release a new Gemini model today [Unverified]: @OfficialLoganK posted a single word, Gemini (5,279 likes); @testingcatalog cited word that a new Gemini model is coming today, most likely Gemini 4 Carbon (778 likes). As of about 22:15 Beijing time on Nov 10, no official Google release had appeared. (@OfficialLoganK · @testingcatalog)
• Codex edition of the ModRetro Chromatic goes on sale at OpenAI Supply Co. for $200: The ModRetro Chromatic plugin in Codex lets you design levels, animate sprites and play-test on the handheld. The product page lists it at $200 with one programmable cartridge, shipping only to the US, Canada, Australia, New Zealand, Japan, the UK and the EU; at the time of writing the page showed it awaiting restock, expected Nov 19. (@OpenAIDevs · Supply Co. product page, 3,127 likes)
• OpenTable says 190,000 people booked through its ChatGPT integration in September [Secondhand account]: Greg Brockman (@gdb) quoted the figure and said connecting to OpenAI can bring a lot of business; the quoted post is from @juokaz (posted Nov 9), who said the figure came from OpenTable’s CFO and that he was double-checking it himself. No official OpenTable source was found. (@gdb, 788 likes)
• Fireworks AI security incident affects Notion; fewer than 1% of workspaces may be affected [Company claim]: Notion says Fireworks notified it on Nov 6 that customer data may have been accessed and confirmed it on Nov 8. The data involved is Notion AI conversation logs used for the Insights feature, stripped of user and workspace identifiers and excluding tool call results. Notion disconnected from Fireworks on Nov 3 and has disabled Insights, and it advises users who pasted passwords or keys into Notion AI to rotate them immediately. (Notion Help Center · HN, 92 points)
• Other high-engagement posts on X: Musk said Starlink’s software is also well suited to training superintelligence (@elonmusk, 39,679 likes) [Opinion]; Ethan Mollick, after early access to Fable 5.2, said it crosses a line of sorts toward jagged AGI (@emollick, 4,051 likes) [Opinion]; OpenCode said MiniMax M3.1 Flash is free in OpenCode, which the Zen docs list as free for a limited time (@opencode, 3,505 likes); Theo used a chart he made to show the shifting usage share of Codex and Claude in T3 Code, saying Opus 5.5 dominates (@theo, 2,044 likes) [Personal test]; SemiAnalysis said Fable 5.2 tops GDPx at 16.4% (@SemiAnalysis_, 834 likes) [Self-reported by the leaderboard].

---

About this column and how we verify

1. AI-assisted: AI Daily Brief is compiled with AI assistance; this issue was produced with Grok Bot and covers Nov 10, 2026 00:00 to about 22:15 Beijing time.
2. Every post checked: Each quoted post was read individually to confirm it exists, the author matches and it was posted within this issue’s time range, and every quoted post was cross-checked a second time through the X API. Like and view counts were read at about 22:15 Beijing time on Nov 10; GitHub stars and HN points are also as of that time and will keep changing. Figures such as prices, benchmark scores and funding amounts were checked against official pages or major media.
3. X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts come from the original posts, official pages and media reports, and the primary source wins in any conflict.
4. Items involving Grok Bot: The brief item “Musk: Grok 4.8 pretraining is done and it’s in RL” concerns the Grok model family; it was included on engagement alone, under the same standards as everything else.
5. Labels: [Company claim] = from a company announcement or official account, not independently reproduced; [Self-reported by the leaderboard] = from the benchmark operator; [Media report] = from media reports, details may differ from official accounts; [Media report, not officially confirmed] = reported by media, not confirmed by the company involved; [Unverified] = circulating without official first-hand confirmation; [Secondhand account] = relaying someone else’s claim; [First-person account] = the person’s own account; [Estimate] = a figure the author estimated from public data; [Personal test] = an individual’s own test or tally; [Opinion] = personal judgment, not a statement of fact.