Anthropic’s Fable 5.2 and Mythos 5.2, a Fireworks breach, and Meta’s cheapest model
Compact Conversations for 2026-11-09: 6 AI stories, ai news worth knowing in just 5 minutes.
[Audio embed placeholder]
The Lead: Anthropic releases Fable 5.2 and Mythos 5.2
Anthropic announced Claude Fable 5.2 and Mythos 5.2, its most capable models for open-ended technical work. They share one underlying model, with outputs up to 30% faster than 5.1, lower pricing at $8 per million input tokens and $40 per million output tokens, and an optional fast mode. Anthropic reports gains in coding, reasoning, and vision, including a thousand-agent team formalizing Fermat’s Last Theorem in 30 hours and a Python-to-Rust migration that starts 26 times faster.
Why it matters: These models are available now on AWS, Google Cloud, and Azure, with Mythos gated behind verification programs, giving enterprise teams a new tier of capability for ambitious technical work.
Source: Anthropic
Number to Know: 90% of Anthropic’s codebase changes written by Claude
Anthropic says over 90% of the changes merged into its own core codebase are now written by Claude, with human guidance. This figure comes from the same release as today’s lead story and reflects the company’s own workflow.
Why it matters: It’s a concrete data point on how AI-assisted development is being adopted inside an AI lab itself, not just in customer benchmarks.
Source: Anthropic
The Feed
Meta releases Muse Ember, its smallest and most cost-efficient model
Meta’s Superintelligence Lab released Muse Ember, its lightest and most cost-efficient AI model yet, designed for repetitive agent tasks like tool calls, coding subtasks, routing, and document parsing. At $0.10 per million input tokens, it undercuts Anthropic’s Haiku 5.5 and OpenAI’s GPT-6 Luna, and handles text, images, and audio in one model.
Why it matters: Ember is positioned as a volume model that pairs with larger models like Muse Wonder, potentially lowering costs for high-volume agentic workflows.
Source: TechCrunch
Fireworks breach exposes customer datasets and model weights
AI inference provider Fireworks says hackers accessed some customers’ datasets, incomplete model weights, and billing information over the past month. Two customers told Forbes they have substantially moved their business off Fireworks’ systems, with one citing a series of problems.
Why it matters: Inference providers run customer models on shared infrastructure, so a breach can expose sensitive data flowing through those models, raising trust and security concerns for enterprises.
Source: Forbes
Google ships four TPU variants in one year
Google added two new eighth-generation TPU systems: the TPU 8tx for confidential computing (data encrypted during processing) and the TPU 8i Max with dual chips on a single package for double compute and memory. Google now counts nearly 100 external TPU customers.
Why it matters: Google’s expanding TPU lineup and growing external customer base signal increasing competition with Nvidia in the AI inference hardware market.
Source: The Decoder
EU draft Digital Fairness Act targets addictive features and AI relationship simulators
A draft of the EU’s Digital Fairness Act would require companies to let adult users disable features that encourage compulsive use, and would target AI systems that simulate personal relationships to keep users interacting.
Why it matters: If enacted, this regulation could force changes in how social media platforms and AI chatbots are designed, affecting companies operating in the EU.
Source: Bloomberg
One Thing to Try
A developer spent 15 weeks building a game called Nightshoal using Claude Opus in the Godot engine, posting progress every Monday on Reddit. The key takeaway is the cadence: pick one small, ongoing project and run it with an AI assistant on a fixed weekly loop. Each week, pick one visible slice of work, build it, and post an honest update—even if the week was mostly polish. The developer found that when commenters asked for more gameplay, the weekly writeup made it obvious what had slipped, and the next week’s plan wrote itself. It only works if updates are honest about what didn’t get done.
Sources
- Anthropic releases Mythos 5.2 and Fable 5.2, saying they are “best for the most ambitious, open-ended technical work”, with outputs up to - Anthropic
- US federal prosecutors say businesses have lost $129M+ to scam calls using AI-generated voices to mimic company executives in the 18 months - The New York Times
- Meta releases Muse Ember, its smallest and most cost-efficient AI model yet, saying it delivers more intelligence per dollar than Haiku 5.5 - TechCrunch
- AI infrastructure provider Fireworks says hackers accessed some customers’ datasets, incomplete model weights, and billing information over - forbes.com
- Google ships four TPU variants in one year as it ramps up competition with Nvidia - The Decoder
- Draft law: the EU’s Digital Fairness Act would require companies to let adults disable addictive features and will target AI systems that - Bloomberg
- 15 weeks of making game with Claude Opus in Godot - itch.io
Transcript
Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.
Host A: Today’s lead is Anthropic’s new model release. The company announced Claude Fable 5.2 and Claude Mythos 5.2 on Monday, and says these are its models for the most ambitious, open-ended technical work. They share one underlying model, and Anthropic reports gains in coding, reasoning, and vision, with outputs up to thirty percent faster than the 5.1 versions. Pricing dropped too: eight dollars per million input tokens, forty per million output tokens, plus an optional fast mode that speeds generation up to two and a half times for double the price. Host B: [curious] The claimed results are the part people will talk about. During prerelease testing, Anthropic says a thousand-agent team formalized the proof of Fermat’s Last Theorem in thirty hours, a Python-to-Rust service migration produced a service that starts about twenty-six times faster, and one engineer’s prototype operating system experiment involved over a thousand subagents working for eight days. On Anthropic’s own coding benchmark, Fable 5.2 at high effort scored about level with Opus 5.5 at max effort, for roughly half the cost per task. Host A: [thoughtful] Worth noting the sourcing: most of those numbers come from Anthropic’s own evaluations and system card, though external groups like the UK AI Security Institute and METR did test the model before release. Fable 5.2 is available today across AWS, Google Cloud, and Microsoft Azure, and Mythos 5.2 stays gated behind Anthropic’s cyber and life sciences verification programs.
Host B: One number to know today: ninety percent. Anthropic says over ninety percent of the changes merged into its own core codebase are now written by Claude, with human guidance. That’s from the same release as today’s lead, and it’s a vendor figure about the company’s own workflow, so treat it accordingly. Still, it’s a concrete data point on where AI-assisted development is landing inside an AI lab itself, not just in customer benchmarks.
Host A: Meta is going after the cheap end of the model market. TechCrunch reports Meta’s Superintelligence Lab released Muse Ember on Monday, its smallest and most cost-efficient model yet, aimed at what Meta calls the work agents repeat all day: tool calls, coding subtasks, routing, and document parsing. Host B: [conversational] The pitch is cost per task. Ember runs ten cents per million input tokens, well under Anthropic’s Haiku 5.5 and OpenAI’s GPT-6 Luna, and competitive with open-weight options like DeepSeek’s flash-tier model. Meta’s framing is that a bigger model like Muse Wonder plans, and Ember handles the volume. It also handles text, images, and audio in one model. Host A: A security story next, from AI infrastructure. Forbes reports that Fireworks, an AI inference provider, says hackers accessed some customers’ datasets, incomplete model weights, and billing information over the past month. Inference providers run customers’ models on shared infrastructure, so a breach there can expose the data flowing through those models. Host B: [thoughtful] Two customers told Forbes they had substantially moved their business off Fireworks’ systems, with one describing a series of problems. The company hasn’t said how many customers were affected, and Forbes notes the full picture is still developing. Host A: The Decoder reports Google shipped four TPU variants in a single year for the first time, adding two eighth-generation versions: one that keeps data encrypted even during processing, and one that pairs two chips on a single package to double compute and memory in the same rack. Google counts nearly a hundred external TPU customers now. Host B: [lighter] And from Brussels, Bloomberg reports a draft EU Digital Fairness Act that would require companies to let adults switch off features designed to drive compulsive use, and would specifically target AI systems that simulate personal relationships to keep users chatting. It’s a draft, so expect the usual negotiation.
Host A: One Thing to Try comes from a developer who’s spent fifteen weeks building a game called Nightshoal using Claude Opus in the Godot engine, posting progress every Monday on Reddit. The tip isn’t the game, it’s the cadence. Pick one small, ongoing project and run it with an AI assistant on a fixed weekly loop: each week, pick one visible slice of work, build it, and post an update, even when the week was mostly polish. This developer found that when commenters asked for more gameplay, the weekly writeup made it obvious what had slipped, and the next week’s plan basically wrote itself. It only works if the updates are honest about what didn’t get done.
Host A: That’s Compact Conversations for Monday. More AI news tomorrow. Until then, happy prompting.