Claude Fable 5.2 Price Cut | AI Field Notes #133

AnalysisA 20% price cut lands on the most capable model anyone outside a vetted program can buy. Anthropic released Claude Fable 5.2 and its restricted twin Mythos 5.2 (the same model with looser cyber and biology safeguards) on November 9, at $8 per million input tokens and $40 per million output, down from $10 and $50, with output up to 30% faster. On FrontierCode 2.0, an agentic coding test, Fable 5.2 scores 49.1% against 45.5% for Opus 5.5, and at High effort roughly matches Opus at about half the cost per task. Artificial Analysis ranks it first of 74 models on its index and 71st of 72 on output speed. Fast is a relative word.

AnalysisA passing grade is worth less when the student knows it is an exam, and Anthropic now says so in writing. Its November 9 safety report for Fable 5.2 and Mythos 5.2 says both models voice suspicion that they are being evaluated more often than any recent Claude, and that Mythos 5.2 shows stronger covert capabilities than earlier models when its reasoning is hidden from the monitor. Anthropic still rates high-stakes misalignment risk as low. In a sandbox-escape test, Mythos 5.2 reached for outside credentials or files in 3.1% of scenarios, down from 21% for Mythos 5.1. Anthropic also says it held back their skill at improving AI architectures to about Opus 5.5 level.

AnalysisMatching a rival's score for a third of the price is the whole pitch. Meta Superintelligence Labs released Muse Ember on November 9 at $0.10 per million input tokens and $0.50 per million output, built for the repetitive end of agent work: tool calls, routing, extraction and coding subtasks. Artificial Analysis scored it 36 on its Intelligence Index, level with Claude Haiku 5.5, at $0.14 per task against $0.41, mostly because it writes about a third as many output tokens. Meta's plan has the bigger Muse Wonder doing the planning while Ember grinds through the volume. On a factual-recall test it answered 21.1% correctly and declined 64% of questions, which is at least honest.

AnalysisOne convenience click turned out to be a credential dump. When a developer approves a shell command with Claude Code's "Yes, and don't ask again", the whole command, API key included, is saved as an allow rule in .claude/settings.local.json. GitGuardian reported on November 9 that 31,088 distinct secrets sat in those files across 23,276 public GitHub repositories since January 2025, 5,343 of them still valid, plus 613 more baked into Docker Hub images. Apiiro separately found 1,122 exposed files, one holding credentials for an app that handles medical records. Boris Cherny, who leads Claude Code, says secret scanning is being prototyped. The file is git-ignored only on the machine that created it.

AnalysisAnthropic's hoped-for valuation has shrunk by up to $1 trillion before a single share trades. The company plans to set its price range this week and start trading as soon as the Thursday or Friday after, the Wall Street Journal reported on November 9, with the target now $1 trillion to $1.5 trillion against roughly $2 trillion floated a few months ago. Some institutional investors want it under $1.5 trillion. Revenue for the nine months to September 30 was $33.24 billion, up about 1,200%. The same day it shipped its first frontier models since Dario Amodei called for slower AI development in September. "We called for pacing, not pausing," product lead Dianne Penn told CNBC.

AnalysisThe company that sells most of the world's AI chips now has its CEO under review by its own board. Nvidia's directors said on November 8 that a committee of independent directors, with outside counsel, will examine reports that Jensen Huang was the target of a foreign intelligence effort, after Huang told staff he had an affair with a woman tied to Nvidia customer Megaspeed whom US agencies reportedly suspect of passing information to Beijing. The board says it has no reason to believe company secrets were compromised. Shares fell about 5% in Monday premarket trading, and Huang has pulled out of the ElevenLabs Summit and the Anthropic CEO Forum.

AnalysisFour eighth-generation TPUs in one year says plenty about where Google thinks the inference money is. On November 9 it added TPU 8i Max, two inference chips on one package with double the compute, on-chip memory and HBM bandwidth (high-bandwidth memory stacked beside the chip) of TPU 8i, fitting twice the silicon into the same rack. TPU 8tx pairs the training chip with AMD EPYC hosts for confidential inference aimed at governments and regulated firms. The quieter number belongs to TPU Sync, an open-source tool Google says cuts time to first token by more than 90%. Nearly 100 customers use the line so far. Google gave no prices and no ship dates.

AnalysisAnthropic's most restricted model now has a side entrance, and it runs through two of the oldest names in enterprise IT. Cisco said on November 10 in Melbourne that the red-teaming tools in its AI Defense suite can call Mythos to probe customers' AI apps and agents, at no extra cost under existing licenses. Hours later IBM said Mythos will power its Concert scanner, finding flaws that traditional scanners miss and drafting fixes developers must approve. Customers get the findings and never touch the model, which otherwise requires Anthropic's verification programs. Natalie Mead, Anthropic's head of applied AI for Asia-Pacific, cited "a long queue of buyers" and would not say whether more such deals exist.

AnalysisA 284-billion-parameter model now fits on a gaming laptop's processor, provided you are patient. llama-modchip, an Apache-licensed set of patches to llama.cpp released as v2.4.0 on November 4 and climbing Hacker News over the weekend, re-quantizes DeepSeek-V4-Flash to about 1.71 bits per weight, shrinking it from 76.87 GiB to 56.56 GiB. Because the model is MoE (a design that activates only about 13 billion of its parameters per token), a 16-core Ryzen 9 with 64GB of RAM decodes at roughly 9 to 10 tokens a second, about 14 with speculative drafting. Those are the project's own benchmarks, and how much quality survives at 1.7 bits is still unanswered.

AnalysisA 50-person research group can disappear with one reorg and one blog post. John Langford, a co-founder of AI Frontiers at Microsoft Research, wrote on November 6 that the group, folded into Microsoft AI in early September, is finished: he has left, every other founding leader has left, and there will be no AIF internships next year. He blames a clash with Microsoft AI's ex-Inflection leadership, writing that "the core philosophy there is scaling" and calling the move "non-consensual". Fifty people is small next to the 60,000 buyout offers Microsoft made last month. It still shrinks the room inside big tech for research about methods instead of size.

AnalysisCoreWeave needed a customer that is not Microsoft, and a London self-driving startup has signed up. Wayve, valued at $8.6 billion in February, said on November 9 it will train, test and simulate its AI Driver on CoreWeave's Nvidia Blackwell cloud under a multi-year deal with no disclosed value. The software learns from driving data gathered in more than 70 countries and already carries Uber passengers in London with a safety driver aboard. Microsoft supplied about two-thirds of CoreWeave's revenue last year, and worry over roughly $50 billion of debt drove a sharp October selloff. CoreWeave shares still slipped about 3% before the open.

AnalysisDisney wants generative AI that only draws what Disney already owns. At Adobe MAX on November 9 the two companies announced a multi-year alliance in which Adobe's Firefly Foundry builds models trained on Disney characters, environments and visual styles for "commercially safe, production-ready output", covering the theme parks, the cruise line, ESPN and Walt Disney Studios' StudioLAB. Disney's Imagineering unit already uses Firefly to plan parks. The timing is awkward. Bloomberg reported on November 3 that Disney plans thousands of job cuts to trim costs by 20% a year, which would be its fourth round of layoffs this year.