A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 10 November 2026 · Issue 9 · 4 min read
Fable 5.2 lands 20% cheaper as Anthropic's IPO week opens; Meta's ten-cent agent model; the $655K revenue-per-head bar
The story that matters most
Anthropic · yesterday
Claude Fable 5.2
Anthropic released Claude Fable 5.2 and Mythos 5.2 yesterday — the same model sold two ways: Fable carries the cyber and biology safeguards and is open to everyone, while Mythos drops them for verified organisations.
Any company running agents on Claude gets an immediate margin lift, and anyone selling frontier coding as a product just watched the bar move. Anthropic still points everyday work to Opus 5.5, so treat Fable as the escalations tier and re-run model-cost assumptions in your companies' 2027 plans. Anthropic's own table puts Fable 5.2 at 49.1% on FrontierCode 2.0, ahead of Opus 5.5 at 45.5%; in one internal run, a 1,000-agent swarm wrote a bootable operating system in eight days.
Rounds and exits
Spines — An agentic AI book-publishing platform, raised a $29M Series A-1 led by Zeev Ventures plus a $35M facility from Twelve, on a $50M run rate.
Coframe — An AI platform that endlessly generates and tests website variants, raised a $25M Series A co-led by Khosla Ventures and Hanabi Capital; $35M in total.
MOVIN — Seoul-based markerless motion capture using LiDAR, raised a KRW 10 billion Series A led by IMM Investment; 70% of its revenue is already from abroad.
Ada — Is being taken over by Dialpad in an all-share deal reported by SaaStr at about US$430M, down from a $1.2B pandemic valuation; no shareholder receives cash.
Anthropic IPO — Is due to fix an IPO valuation range this week and begin trading late next week, reported by the WSJ, via TechCrunch; some investors want in below $1.5T.
Creandum Fund VIII — The European seed firm's biggest close yet at $750M, announced alongside its first San Francisco office the week Sequoia shut London.
US Series A — Startups raised $42B through October, up about 60% year over year while deal counts held near recent years' pace, Crunchbase News reports.
What the AI labs and big platforms did
Muse Ember — Meta's smallest model, out November 9, targets the repetitive steps inside agents — tool calls, routing, long-document parsing — rather than the planning layer, and lists at $0.10/$0.50 per million input/output tokens with a 1M-token context window. Meta claims it matches Claude Haiku 5.5's best Artificial Analysis score at about a third of the cost. This is the price floor dropping for agent infrastructure: it squeezes startups reselling cheap orchestration, and forces anyone with high-volume inference to re-cost their stack. Meta's own table still has Haiku ahead on real coding: 34.9% versus 27.2% on FrontierCode 2.0-AA. Ember pairs with the larger Muse Wonder — Meta's line is that Wonder plans and Ember handles the volume. Meta
TPU 8tx and TPU 8i Max — Two more eighth-generation TPU systems announced November 9: 8tx keeps data encrypted even while in use — built for regulated enterprises and governments — and cuts latency up to 2x versus 8t, while 8i Max packages two 8i chips to double compute and memory bandwidth for long-context agent work. Google says TPUs now serve nearly 100 customers. Confidential inference capacity opens regulated verticals — health, finance, government — where frontier models previously could not touch sensitive data, and gives inference-heavy startups a credible Nvidia alternative. TorchTPU, open-sourced last month, lets PyTorch code run on TPUs without rewrites. Google
For your portfolio
SaaStr · yesterday
Revenue per head is now a board question, not a footnote
SaaStr's Jason Lemkin assembled the 2026 picture on November 9, and the spread inside B2B is now a chasm. By ICONIQ's count, the ordinary B2B company earns about $237K of ARR for each person it employs; the fastest growers at $100M-plus ARR are near $655K each and still compounding, while Palantir, Anthropic and AppLovin have all crossed $1.5M.
The old planning ratio — one hire per $200-300K of new revenue — still governs most hiring plans, but the leaders now add a person per $1.75M to $9M of revenue. Lemkin's cut of the data: below $300K per head while growing under 30% will draw board questions in late 2026. Cloudflare's ~20% May cut lifted its figure from roughly $496K to $690K while growth accelerated. Ask every founder you back this quarter: what is revenue per employee today, and which function — G&A first, where Cloudflare cut — gets thinned to reach $650K before the 2027 plan is written?
Aphyr · yesterday
Harvest
Takeaway: $13.32 to $133.98 a month
Kyle Kingsbury, writing as Aphyr on November 9, chronicled what has happened to a beloved freelancer invoicing tool since Bending Spoons bought it in 2025: his bill went from $13.32 to $133.98 a month when everyone was moved onto an enterprise-style usage plan, the cheap flex tier adds a 2% skim on gross revenue on top of card fees, and support mail is now answered by a language model rather than staff.
This is the buy-and-squeeze rollup playbook in the wild, and it lands on two portfolios at once: SMB software you back now competes with repriced incumbents whose users are actively churning, and any founder weighing an all-stock sale to a private rollup should read what happens to the customers and the brand afterwards. Ask your SMB-facing founders: if the incumbent gets repriced 10x overnight, are your pricing and your human support the wedge — and would you take a rollup's cash today?
“I did not anticipate the speed at which Bending Spoons would drive a perfectly good company into the ground.”