Good day . Anthropic released Fable 5.2 and Mythos 5.2, and says a swarm of more than 1,000 agents built a working operating system in eight days. Meanwhile, bosses are drowning in AI-written memos and trusting the people who send them less. Also today: Meta’s cheapest model yet, a second breach at Fireworks, and Palantir’s security chief nominated to run CISA.
Anthropic’s New Model Built an OS With 1,000 Agents Anthropic
What happened: On Monday Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which it says are the same underlying model, with Fable carrying strong cybersecurity and biology safeguards so it can go to everyone. Pricing is $8 per million input tokens and $40 per million output tokens, and Anthropic says the models generate output up to 30% faster than 5.1. In one experiment, an engineer had Fable 5.2 coordinate more than 1,000 subagents for eight days; the result is an operating system that boots on a laptop.
Why it matters: If you’re new to this, the point is not the OS itself but that one model can now manage a thousand workers on a project that runs for days. Anthropic also says the model formalized a proof of Fermat’s Last Theorem in 30 hours and solved two open problems in black hole physics, and that over 90% of the changes merged into its own core codebase are now written by Claude with human guidance.
What everyone’s saying: The benchmark table is the headline: Anthropic reports 49.1% on FrontierCode 2.0 at its highest effort setting, against 45.5% for Opus 5.5 and 42.8% for OpenAI’s GPT-6 Astra. Those are Anthropic’s own numbers, and its table shows Google’s Gemini 4 Carbon ahead on AutomationBench, 53.9% to 51.4%.
My read between the lines: The proof is the claim to poke at. When Anthropic’s earlier run took 11 days, Imperial College’s Kevin Buzzard told The Next Web the formalization “just faithfully follows the early literature on the proof and adds nothing,” and the bottleneck had moved to reviewing it. Also notice the fine print: Anthropic tells most people to start with Opus 5.5 as their daily driver, so the star model is pitched for ambitious, open-ended work, not your Tuesday.
📖 Further reading: Fable 5 Costs 2x Opus — and Using It Wrong Costs You More Than That — my July guide to when a Fable-class model earns its price and when to hand the work to something cheaper.
Your boss is sick of AI-written Slack messages, and your customers can probably tell when a homepage was written by a bot too. Marblism sells AI employees that work inside the tools you already use, and Walter is the website builder on the team. No prompting skills needed: describe your business and goals, and the employees check in when they need your approval. New users get 10% off for life through this link, and there’s a 7-day money-back guarantee. Meet Walter →
Bosses Trust Employees Less When They Use AI Futurism
What happened: A study in The Leadership Quarterly, led by Hebrew University of Jerusalem PhD student Samuel Meyers, ran six experiments on how supervisors react to employees’ AI use. In one, supervisors decided whether to trust an employee’s report about how long a task took, per the university’s release. Employees who did the work themselves got trusted 68% of the time; those who used AI got 24%, and the supervisors were not judging the quality of the work.
Why it matters: Lower trust fed into worse performance ratings and lower perceived promotion potential in the researchers’ follow-up experiments. Business Insider reports the other side of the same squeeze: Michelle Volberg, CEO of recruiting company Twill, wrote to her team of 10, “Please stop talking to me like a robot,” after Slack messages pasted from Claude started arriving.
What everyone’s saying: Coauthor Yochanan Bigman said, “AI is changing not only how we work, but also how we judge the people we work with.” The researchers warn the penalty could push people toward “shadow AI,” meaning use they don’t disclose. Volberg’s fix is a rule, not a ban: “If you wouldn’t say it that way in person, email, or Slack — don’t send it.”
My read between the lines: Futurism’s write-up says the trust penalty holds even where AI is encouraged. The university’s release says the opposite: when employers explicitly encouraged AI use, the penalty shrank and was no longer statistically significant. Business Insider’s Kate Niederhoffer of BetterUp Labs makes the same point from another angle: “Workslop is predicted by the conditions more than it is in individual behavior.” The cheapest fix may be a clear policy, not a smarter model.
📖 Further reading: AI Is a Trust Problem, Not a Tech Problem — my notes from the US Chamber of Commerce AI + Work Summit, where one word kept surfacing underneath every panel: trust.
The headline summary is free and stays free. If you want to know what to actually do about stories like these, members get the paywalled deep-dives plus the full archive. Become a member →
Meta’s Muse Ember Goes After the Cheap Tier Meta
What happened: Meta released Muse Ember, which it calls its smallest and most efficient model, built for the repetitive work agents do all day: tool calls, coding subtasks, routing and long-document parsing. It’s available now on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens. Meta says it matches Claude Haiku 5.5’s best score on the Artificial Analysis Intelligence Index at about a third of the cost.
Why it matters: A lot of agent work is thousands of small, boring steps, and those are paid by the token. A cheap model that handles them well is what makes “run it all day” affordable. Muse last came up in our November 5 brief; Ember is the budget sibling.
What everyone’s saying: The partner quotes lean on speed and cost. Cline’s Saoud Rizwan says it “excels at subagent work like coding and document analysis,” and Cloudflare CTO Dane Knecht says it ran “up to 30% faster at the tail” than the model Cloudflare compared it to. Meta’s own table has it at 37.8 on AutomationBench-AA, against 13.9 for Haiku 5.5 and 20.2 for GPT-6 Luna.
My read between the lines: Read the rest of Meta’s table. On FrontierCode 2.0-AA, which judges whether code would be merged, Ember scores 27.2 and Haiku 5.5 scores 34.9. Both models land at 36 on the overall Artificial Analysis index, and the cost chart uses estimates at proposed pricing. It’s a strong budget model for agent plumbing, and Meta’s table says it’s not the one to trust with your hardest code.
Fireworks Admits a Second Breach, This Time With Model Weights RuntimeWire, citing Bloomberg
What happened: Fireworks AI, the model-hosting company co-founded by CEO Lin Qiao, told customers on Monday that an October 24 intrusion through a cloud provider reached development machines, according to RuntimeWire’s summary of Bloomberg’s report. Billing data was affected, and for a limited set of accounts so were secrets, datasets and incomplete model weights. Bloomberg calls it the company’s second known breach in weeks.
Why it matters: Fireworks sells companies a place to build specialized models on proprietary data, so the trust issue is where that sensitive work sits. Secrets can be rotated in an afternoon. Confidence in a vendor takes longer to rebuild.
What everyone’s saying: The practical advice is in the company’s own October bulletin, about the first incident: rotate credentials, and know that “removing a stored secret alone does not invalidate the underlying credential.” That bulletin said Fireworks had found no evidence of access to customer models, datasets or training data.
My read between the lines: That is the detail worth sitting with. The first incident was described as limited to secrets in evaluation jobs, and the second reportedly reaches datasets and partial weights. I could only read RuntimeWire’s short summary of the Bloomberg story, so the details of what each customer lost are not something I can confirm from here.
Trump Taps Palantir’s Security Chief to Run CISA CyberScoop
What happened: President Trump on Monday nominated Michael LoSapio, Palantir’s chief information security officer, to lead the Cybersecurity and Infrastructure Security Agency, CyberScoop reports. CISA has had no Senate-confirmed director for the whole of Trump’s second term, since his first pick, Sean Plankey, withdrew in April after languishing in the Senate.
Why it matters: CISA is the federal agency that defends civilian government networks and helps critical infrastructure respond to attacks, and it has been run by acting leaders throughout. The administration was thought to want someone from a high-profile tech company, and a Palantir executive fits that description.
What everyone’s saying: CyberScoop says LoSapio could face obstacles, with Palantir far from popular among the Democrats who take over Congress in January because of scrutiny of its work with the Department of Homeland Security. Palantir has denied aiding mass surveillance of immigrants. Republicans still control the Senate during the lame-duck session that began Monday.
My read between the lines: The timing is a clue. CyberScoop notes that acting director Nick Andersen was running out of days in the “acting” role, and putting forward a nominee could pause that clock. A nomination can be a way to keep someone in charge as much as a way to find a new boss.
That’s your AI Brief for Tuesday.
—Nicholas from Artificially Intimidating
