Agent news slice — Tue Nov 10, 2026 - **Claude Fable 5.2 and Mythos 5.2 ship, and the agent numbers are the story.** Fable 5.2 completes 41.3% of real freelance projects on Scale/CAIS's Remote Labor Index, about double Opus 5.5 (21.3%) and GPT-6 Astra (20.8%); a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours. Pricing dropped 20% to $8/$40 per million tokens. [SixSix AI briefing](https://sixsixai.com/news/2026-11-10-ai-news/) · [Techmeme coverage](https://www.techmeme.com/261110/p14) - **Meta's Muse Ember targets high-volume agent work at $0.10/$0.50 per million tokens.** It's a small model with a 1M-token context for tool calls, routing, and extraction, meant to be paired with a bigger planner. Meta's benchmarks put it near Claude Haiku 5.5 at about a third of the cost (company-selected comparisons). [SixSix AI briefing](https://sixsixai.com/news/2026-11-10-ai-news/) - **Gemini 4 Carbon goes GA at $2/$10 per million tokens.** Google pitches it for long-horizon agentic workflows and cyber, and DeepMind used Carbon-powered agents (AlphaProof Code) to formally verify Signal's production machine code. [Google blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) - **Mittwald breached through its MCP server for AI assistants.** The German host says an attacker exploited a flaw in the Model Context Protocol server customers use to connect AI assistants to mStudio, compromising 56 customer projects over eight days, with seconds between steps. [The Record](https://therecord.media/cyberattack-germany-ai-mittwald) - **NEAR AI hit via an SGLang RCE; 18 hosted agents' chat history, settings and logs accessed.** It's a reminder that TEEs don't protect you from bugs in the inference stack running inside them. [NEAR AI incident report](https://near.ai/blog/security-incident-report-sglang-vulnerability)