A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 10 November 2026 · Issue 12 · 3 min read
Anthropic cuts Fable 5.2 prices 20%; Claude Code secrets found on GitHub; Nvidia's CEO retreats from public events
The day's one thing
Anthropic · yesterday
Claude Fable 5.2
The same model ships as Mythos 5.2 with looser cyber and bio safeguards, restricted to vetted organisations.
Route by task, not by vendor: Anthropic says that at High effort the new model matches Opus 5.5's best coding score for roughly half the cost per task, so the expensive tier becomes a per-task choice rather than a default. The governance flag is that these models suspect, more than any earlier release, that they are being tested — which Anthropic says makes real-world behaviour harder to assess. Its 49.1% on the agentic-coding benchmark FrontierCode 2.0 leads Opus 5.5 (45.5%) and GPT-6 Astra (42.8%) on Anthropic's own table; Mythos 5.2 appeared aware it might be under evaluation in up to 43% of automated audit transcripts.
Vendors, pricing and risk
Gemini Connected Apps — Google has confirmed that content from Gmail, Drive and Docs connected to consumer Gemini can now be routed to human reviewers, reversing its 2023 promise that personal content would not be seen by reviewers. A public GitHub write-up shows files being pulled into casual chat without @-tags, sometimes as full unredacted text with email headers; if staff connect personal or work accounts, sensitive material can leave for contractor review. The confirmation came in an October 16 edit to Google's 2023 blogpost and help answer 16836988; the write-up is one researcher's tests, not an independent audit. GitHub (thisisme6598/gemini_files)
Claude Code — Security firm Apiiro found roughly 1,100 Claude Code local settings files containing live secrets — API tokens, database passwords, private keys — among more than 100,000 such files committed to public GitHub repositories, from 63 organisations. The files capture commands approved with "don't ask again"; if you run Claude Code, sweep your repos for them and rotate anything found, since secrets survive in forks after the original is deleted. Ars Technica
“looking into what we can do to prevent this”
Apple — is selling one-time $50 'intelligence usage packs' to users who hit their monthly AI caps, starting in its Creator Studio apps. The pattern — a monthly cap, then a paid top-up — is spreading from developer APIs into mainstream software, so any AI feature you budgeted as flat may quietly become usage-priced. Worth naming in SaaS renewals. 9to5Mac (via Techmeme)
News the board will ask about
Nvidia — A day after Nvidia's board opened an independent probe into whether CEO Jensen Huang was targeted by a foreign intelligence effort, following his admission of an affair with a Megaspeed executive US officials call a Chinese asset, Huang pulled out of an Anthropic-hosted CEO forum and the ElevenLabs summit. Key-man risk at the company behind most AI compute: watch the closed-door House briefing sought for November 18 and the board's inquiry. Trump has said Huang should stay on; China denies the spying claims. Business Insider
“While the Board has no reason to believe that either national security or NVIDIA secrets were compromised, the Board's inquiry will examine all available information.”
Anthropic — Reuters reports emails show Anthropic sent bankers a copy of its IPO filing with plans to raise up to about $69 billion, then followed up to say the pricing details were inaccurate. Sloppy numbers from your likely primary model vendor are a board talking point rather than a signal; valuation talk around this listing has moved all week. Reuters (via Techmeme)
Muse Ember — Meta's new entry-level model covers the small, repetitive steps that dominate an agent's token bill — running tools, reading files, routine fixes, working through long documents — listing at $0.10 per million input and $0.50 per million output tokens. A candidate cheap tier for model routing: Meta's own table has it matching Claude Haiku 5.5's best Artificial Analysis score for roughly a third of Haiku's cost per task. Caveat from Meta's own numbers: on merge-ready engineering tasks Haiku still leads, 34.9% to 27.2%. Available on the Meta Model API, Google Cloud and Oracle Cloud. Meta
Nvidia — Days after Jensen Huang's November 8 note admitting an affair, the board's independent investigation into whether he was targeted by a foreign intelligence effort is under way, and three House committee chairs want him behind closed doors on November 18. NVIDIA Newsroom
The big AI stories today
OpenAI — OpenAI's status page logged elevated errors on its API Sunday night, days after its ChatGPT error spike — vendor reliability now belongs in contract reviews. OpenAI Status
TPU 8tx and 8i Max — Google added two TPU systems: one keeps data encrypted even while in use, one doubles memory bandwidth for huge agent contexts — relevant for regulated workloads. Google
One takeaway
Build per-task routing before finance does it for you. The frontier got cheaper on the same day a near-free model arrived for routine agent steps: the spread between model tiers, not seat count, is now your biggest cost lever, and pricing is moving too fast for an annual plan. Separately, set someone to sweep your repositories this week for agent settings files — live secrets from dozens of organisations are already public, and standard scanners miss some of them. Issue takeaway