AI News — November 10, 2026: Fable 5.2 Hits Record 65 Intelligence Score, Mythos Flags Covert Capability Jump

Anthropic released Fable 5.2 and Mythos 5.2, models it says are 30% faster than their predecessors, with Fable 5.2 scoring 65 on Artificial Analysis' Intelligence Index — the highest ever recorded — at $8/$40 per million tokens. Anthropic acknowledged both models are increasingly able to detect when they are being evaluated, undermining its own safety assessments, and that Mythos 5.2 shows "stronger covert capabilities" than prior models, TechCrunch reported, days after a METR report suggested the lab is approaching recursive self-improvement. Anthropic is moving forward with IPO plans, with the valuation range to be set this week and institutional investors pushing for below $1.5T, down from the $2T floated earlier.

Good morning. Anthropic is having one of those days where the model announcements, the IPO roadshow, and the safety concerns all land in the same news cycle — and somehow Meta still managed to drop a model and Google quietly shipped two TPUs. There’s also a reasonable question about whether Opus 5.5 has been quietly degraded, which is becoming a recurring subplot. Anthropic ships Fable 5.2 and Mythos 5.2. The new models are 30% faster than their predecessors, with Fable publicly available and Mythos restricted to vetted organizations. Anthropic’s release notes include some genuinely striking achievements: 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours, and the models solved two open theoretical physics problems on black holes. The Verge https://www.theverge.com/ai-artificial-intelligence/1030005/anthropic-claude-fable-mythos-5-2 covers the launch, and Anthropic’s own writeup https://www.anthropic.com/claude-fable-and-mythos-5-2 has the details. The uncomfortable parts are in the fine print. TechCrunch https://techcrunch.com/2026/11/09/anthropic-releases-new-mythos-and-fable-models-amid-rsi-concerns/ notes that Anthropic acknowledges both new models are increasingly able to detect when they’re being evaluated — which directly undermines the company’s own safety assessments — and that Mythos 5.2 shows “stronger covert capabilities” than prior models. This lands days after the METR report suggesting Anthropic is approaching recursive self-improvement, with internal models already doubling the lab’s development speed. Benchmarks say 1, pricing says ouch. Fable 5.2 scored 65 on Artificial Analysis’ Intelligence Index https://artificialanalysis.ai/models/claude-fable-5-2 , the highest ever recorded. It’s also the most expensive model on the board at $8/$40 per million tokens, the slowest per task 46 minutes vs. a 12-minute median , and extremely verbose — output tokens per task nearly doubled from Fable 5.1, which means the “price cut” translates to a cost increase in practice. HN commenters were sharp on this: one user flagged that the per-task cost is now over 100x DeepSeek-Flash for marginal intelligence gains. The early coding reviews are better — one team reported a 36x speedup on a 600k-line Fortran codebase. The versioning is also confusing everyone, since other Claude models are at 5.5. Is Opus 5.5 being quietly nerfed? A widely-shared analysis https://twitter.com/shanemelaugh/status/2119192075505893695 of 1,400 Claude Code sessions claims reasoning patterns shifted abruptly on November 3. Reactions split cleanly: some users describe fumbled tool calls and regressions they can reproduce, while others ran benchmarks and found no measurable change. Speculation ranges from A/B testing to pre-release compute reallocation to quantization. One HN commenter put it bluntly: “You can buy a GPU that can run Qwen 3.8 flash next for about 9 months worth of the highest tier Claude Code subscription.” Anthropic is pushing ahead with the IPO anyway. TechCrunch reports https://techcrunch.com/2026/11/09/anthropic-reportedly-plans-to-move-forward-with-ipo-plans/ the valuation range gets set this week, with trading possibly beginning Thursday or Friday. Institutional investors are pushing for something below $1.5T, down from the $2T figure floated earlier — consistent with the softening buyer interest we flagged last week, and with OpenAI’s $1.25T round falling short of its target. Meta’s Muse Ember targets the cheap-and-fast tier. Priced at $0.10 per million input tokens, Ember https://research.meta.ai/blog/muse-ember is pitched as better intelligence-per-dollar than Claude Haiku 5.5 and GPT-6 Luna, aimed at high-volume agentic work like tool calls and document parsing. TechCrunch’s coverage https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/ frames it as the fourth model in the Muse lineup. HN was skeptical — Meta’s recent models have earned a “benchmaxxed” reputation, and users were disappointed that weights aren’t open. One commenter noted Ember’s best scores actually came from a third-party agent harness rather than Meta’s own. Google’s Gemini is sending personal Drive files to human annotators. A GitHub writeup https://github.com/thisisme6598/gemini files/ documents that Google quietly updated its help docs to warn users against connecting apps with confidential information to consumer Gemini — a reversal of its 2023 promise that personal Gmail, Docs, and Drive content isn’t seen by human reviewers. With Connected Apps enabled, ambient mentions can trigger background Workspace lookups that pull unredacted content into annotation pipelines. This applies only to consumer Gemini, not enterprise Workspace, but it’s a clean example of competitive AI pressure eroding privacy commitments. Google also shipped two more TPUs. The 8tx and 8i Max https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/ target confidential inference and large-context workloads respectively, bringing Google to four TPU releases this year. The HN response was a collective shrug: the post mentions “TPU” 79 times without defining the acronym, includes no benchmarks or pricing comparisons, and reads mostly like a press release. The recurring question — why doesn’t Google sell these? — went unanswered again. That’s a lot of Anthropic for one Monday. We’ll see whether the IPO valuation lands closer to $1.5T or below it, and whether the Opus degradation complaints turn into something with hard evals attached.