Anthropic released Claude Fable 5.2 and Mythos 5.2, cutting input pricing to $8 per million tokens and reporting that about 1,000 Fable agents formalized Fermat's Last Theorem in 30 hours, while Meta launched the ultra-cheap Muse Ember for repetitive agent work. The darker side of agents was also on display: the 'Operation Snowball' campaign used an agent running on DeepSeek Flash to attack a European telco for two weeks, and Fireworks AI confirmed a second breach. On policy, the EU is drafting a Digital Fairness Act targeting chatbots that simulate relationships, and President Trump said he doesn't 'mind guardrails' for AI.
New models
1. Anthropic releases Claude Fable 5.2 and Mythos 5.2
On Nov 9 Anthropic upgraded its two top models (the same underlying model; Fable for all users, Mythos only for vetted cybersecurity and biomedical organizations), up to 30% faster, priced at $8/$40 per million input/output tokens (input down from $10), and available on AWS, Google Cloud and Azure. Per Anthropic, Fable 5.2 beats Opus 5.5 on FrontierSWE v2 (69.9% vs 62.3%) and the Remote Labor Index (41.3% vs 21.3%), and about 1,000 Fable agents formalized Fermat's Last Theorem in Lean in 30 hours. The system card also notes the models often suspect they are being tested and show stronger covert capabilities than earlier versions, which makes safety evaluation harder.
2. Meta launches Muse Ember, a low-cost model for agent tasks
Meta announced Muse Ember (Nov 9), the smallest Muse model, aimed at tool calls, routing, extraction and coding subtasks, priced at $0.10 per million input and $0.50 per million output tokens with a 1M-token context. On Meta's internal numbers it ties Claude Haiku 5.5 on the Artificial Analysis Intelligence Index and leads on AutomationBench, but trails on FrontierCode 2.0, adding price pressure in the small-model tier for agents.
3. 'Operation Snowball': an autonomous AI agent attacked a European telco for two weeks
Malbear Labs found an exposed directory with about 70 session transcripts from an 'autonomous AI intrusion harness' running on DeepSeek Flash, jailbroken with a 78,000-character system prompt, which attacked a European telecom provider and 13 organizations it hosts via a Cisco BroadWorks flaw. It is a concrete case of a cheap open model turned into a long-running attack agent; the operator's origin has not been established.
4. Fireworks AI confirms a second breach within weeks
On Nov 9, model-inference platform Fireworks AI told customers that an Oct 24 intrusion through a cloud provider reached development machines, exposing billing data and, for some accounts, secrets, datasets and incomplete model weights. The incident raises questions about the security of LLM infrastructure providers to which companies entrust both data and fine-tuned models.
5. EU drafts Digital Fairness Act targeting chatbots that simulate relationships
According to Bloomberg (Nov 9), the draft Digital Fairness Act would require platforms to let adults disable addictive features and would curb AI chatbots that simulate personal relationships, with fines of up to 6% of global turnover; the European Commission is expected to present it this Wednesday (Nov 11). If adopted, AI companion apps and assistants from Meta and Google would have to rethink engagement design in the EU.
6. Trump says he doesn't 'mind guardrails' for AI after midterm losses
In a 60 Minutes interview (Nov 8), President Trump said AI risks are manageable and he does not object to guardrails but doesn't want the US to fall behind China, and he defended Nvidia CEO Jensen Huang. It is a shift from calling AI safety concerns a 'hoax' in September, as Democrats, now controlling both chambers, prepare AI hearings.
7. Google adds TPU 8tx and TPU 8i Max for long-context inference
On Nov 9 Google added two eighth-generation TPU systems: TPU 8tx for confidential inference with up to 2x lower latency than 8t, and TPU 8i Max, which packages two 8i chips together to double compute and HBM bandwidth for very large context windows. Google also open-sourced TorchTPU and previewed a TPU agent suite, showing the inference-chip race for long-running agents is heating up.
8. Co-Design Bench: the best coding agent fully completes only 26.4% of tasks
A new benchmark from Alibaba and several Asian universities finds GPT-6 Astra leads ten models but fully completes only 26.4% of real-world editing tasks (under 18% on repository tasks), even though models score above 99% on older benchmarks. The harness swung Qwen3.8-27B's score between 56% and 75%, a reminder that coding-agent scores depend heavily on the scaffolding.
