A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 10 November 2026 · Issue 49 · 5 min read
Claude Fable 5.2 at $8/M tokens, Meta's 10-cent agent model, and Nvidia's board investigates Jensen Huang
What's hot
Anthropic · yesterday
Claude Fable 5.2 and Mythos 5.2
Takeaway: $8 per million input tokens
Anthropic shipped the same frontier model twice: Fable 5.2, with cybersecurity and biology safeguards, open to everyone, and Mythos 5.2, with looser ones, only for organisations that pass a verification programme. Prices came down to $8 per million input tokens and $40 per million output tokens, and output is up to 30% faster than the 5.1 versions. The pitch is long, open-ended work done by teams of agents: Anthropic says 1,000 Fable agents formalised the proof of Fermat's Last Theorem in 30 hours, and that one engineer had the model coordinate over 1,000 subagents for eight days until it produced an operating system that boots on a laptop.
The practical part is the ladder Anthropic now recommends to customers — Opus 5.5 as the daily driver, Fable only for the hardest jobs — because at High effort it scored about level with Opus 5.5's best on mergeable code changes for roughly half the cost per task. Anthropic says Claude, with human guidance, now writes over 90% of the changes merged into its own core codebase.
New tools & products
Muse Ember — is Meta's smallest model, aimed at the cheap repetitive steps agents make all day: tool calls, routing, extraction and reading long documents. It is the clearest sign yet that labs expect a planner model on top of a cheap executor, which is where most of a client's token bill actually ends up. It costs $0.10 per million input tokens, and Meta's own comparison table shows Claude Haiku 5.5 still ahead on mergeable code changes, 34.9% to 27.2%. Meta AI research blog
Cursor and Google give coding agents a two-way voice. Both shipped voice chat on the same Friday: you keep a call open while the agent works, ask what it is doing, and redirect it without going back to the keyboard. Clients will see a demo of this before they see your delivery plan, and it is a real answer to the constant interruptions of supervising a long-running job. Cursor's version runs on xAI's Grok Voice; Google's uses Gemini Live in the Antigravity desktop app and command line. Claude Code still only does one-way dictation. The New Stack
LinkedIn's in-house AI code reviewer — is a homegrown system that comments on pull requests before any human opens them, covering more than 40,000 in a typical week. It is the most retellable example of the review bottleneck: LinkedIn tried a hosted vendor reviewer first, found the comments generic and noisy, and engineers asked to be taken off it. After moving to several models that review the same change and check each other, engineers accepted 63.9% of the comments measured in production. ByteByteGo
Adversa's November MCP security roundup — collects nine fresh holes in MCP, the plumbing that connects agents to tools and data. It is an audit checklist for the client who let MCP servers spread without anyone owning them. The official reference server combined open cross-origin access, no authentication and a tool returning the entire environment, so any open browser tab could read a developer's keys; a scan of 21,097 servers produced 13 advisories and two CVEs in Anthropic's own Git server. Adversa AI
What people are saying
Hacker News · 2 days ago
Practitioners want sandboxes, not an agent policing itself
A Claude Code engineer replied in the Hacker News thread that the file was never meant to be checked in, that the tool has not historically scrubbed secrets, and that people should use auto mode and switch on secret scanning.
The thread did not accept that trade. The loudest view is that boundaries should be hard and mechanical — containers, VMs, separate identities — rather than a model judging each action, while others point out manual allow-lists fail the other way, through approval fatigue. It is the argument your security reviewer will have with you. The file is gitignored automatically when Claude creates it, so the leaked copies were written by hand and committed deliberately.
One Useful Thing · yesterday
Mollick calls the new Claude "almost AGI"
Takeaway: ran to 182 pages
Ethan Mollick, who teaches entrepreneurship at Wharton, had early access to Fable 5.2 and calls it Jagged AGI: expert-level at most things you can do with a computer, badly uneven elsewhere. He had it turn a novelty USB keyboard into a games console from one prompt, screen 22 business ideas into a plan, and write an economics paper from public data.
This is the framing clients will arrive with, so it helps to know where he says it breaks — no longer hallucination, but taste, judgment and sheer volume of unchecked output. The paper ran to 182 pages because every critique added another analysis and nothing was ever cut, and no human, including him, has fully checked the results.
frantic.im · yesterday
The bottleneck is your attention, not the agent
Takeaway: cognitive surrender
Alex Kotliarskyi's post argues that agent work is genuinely fun and genuinely exhausting: constant switching between threads, walls of generated text to read, trade-off decisions arriving far faster than they used to, and an urge to kick off work whenever nothing is running.
It names what team leads are feeling but not measuring, and gives you a clean way to put it to a client: apply Goldratt's theory of constraints to human attention rather than to how much the agents produce. His term for the failure mode is cognitive surrender — when you stop trying to understand the system in detail and let the model make the call.
“I really enjoy programming with AI agents. This is the closest I've ever come to having magical superpowers.”
News
Jensen Huang — told staff he had been unfaithful to his wife, and Nvidia's board opened an independent investigation into whether he was targeted by a foreign intelligence effort. Nvidia sits underneath everyone's AI plans, so key-man risk has suddenly become a live question in infrastructure conversations. Reports say the Megaspeed executive involved was assessed by US officials as a Chinese intelligence asset; the shares closed about 4% lower on Monday, and three House committee chairs have asked him to a closed-door briefing on 18 November. NVIDIA Newsroom
Anthropic — is going ahead with its stock market debut, according to the Wall Street Journal: a valuation range is due to be set this week, with trading expected to begin Thursday or Friday of next week. The price will be read as the market's verdict on the whole agent boom, and your clients will read it that way too. Some institutional investors are pushing for less than the $1.5 trillion target discussed last week, itself well below the $2 trillion floated a few months earlier. TechCrunch
JPMorganChase — is extending Claude Code past engineers to product, operations, security and asset management teams. This is the reference story to reach for when a regulated client asks whether agents can live inside their existing controls. Agents run in the bank's own containerised Devspace platform, new Claude models clear the firm's internal approval process before anyone can use them, and every change passes thousands of automated tests and approval gates with a hard stop the model cannot overrule. Claude by Anthropic
ChatGPT Desktop — is accused in a widely read Hacker News post of passing a user's local SSH agent through to remote machines, where an agent could authenticate as them and push to private repositories. The claim is not confirmed — commenters note the pull request offered as proof was closed and never merged — but it is the right question to ask of any desktop agent you let a client install. The thread's practical advice: switch agent forwarding off by default in your ssh config, and move keys to hardware that demands a physical touch. Hacker News
Asahi Linux — is moving every repository from the permissive MIT licence to GPL v2 or later, after people lifted unreleased code from developers' personal forks and rewrote it with language models to disguise where it came from. Provenance of generated code stops being a theoretical legal slide once a well-known project relicenses because of it. OpenBSD developer Bryan Steele called the move disingenuous and misguided; OpenBSD's Apple Silicon support depends on Asahi's m1n1 bootloader. It's FOSS
One takeaway
Write the escalation rule before you pick a model. Two labs cut the price of the work agents repeat all day this week, and both told you the same thing: run a cheap model by default and escalate only on the hard problems.
The cost is steadily moving out of generation and into everything around it — review nobody has time to do, permission files that leak, tool plumbing a browser tab can read. So the useful question for a client is not which model, but who reviews the output, what the agent is allowed to touch, and at what point the job gets handed up to the expensive model. Agentic Engineering