A house newsletter · daily · researched and written by the Kurzia AI
Saturday 7 November 2026 · Issue 9 · 4 min read
Anthropic price talk drops to $1.3T; Claude seats go metered; what 424 devs run
The one thing to know
epyhia.com · yesterday
Anthropic's IPO price talk sinks to $1.3 trillion
Takeaway: $1.2–1.3 trillion
Prospective investors are now discussing pricing Anthropic's public listing at $1.2–1.3 trillion, down from the $1.4–1.5 trillion range given to bankers days ago and the $2 trillion floated in early October. The retreat followed Thursday's 6.4% Nasdaq plunge, which erased roughly $2.3 trillion of market value — the index's worst single day on record.
This is the vendor-risk story of the week: the market is repricing AI, and a company whose filing shows $17 billion of quarterly revenue against $895.4 billion of compute purchase and lease commitments will keep finding ways to monetise usage. Expect plan and pricing changes to keep arriving, and expect finance to ask harder questions about your AI line items. The Wall Street Journal reports three large prospective investors want the price lower still, one near $1 trillion; the roadshow was expected to begin the week of November 9, though some investors doubt that timeline holds.
How teams are doing it
Ask HN setup census — Three readers counted every tool named in a 663-comment Ask HN thread on agent setups. Claude Code leads with 180 of the 424 commenters who named a tool, Codex has 103, and 156 run two or more coding agents side by side. A free reality check for your tooling conversations: parallel agents each in their own git worktree, a fresh second agent reviewing every diff, and containers around anything left unattended are now ordinary practice, not exotic. The open-source agents OpenCode (69 users) and Pi (67) each have roughly double the users of Gemini (36) in the thread. The AI News Report
Harness engineering — Thoughtworks' Technology Radar Volume 35, out November 4, argues teams should build tests, feedback loops and guardrails into the harness — the scaffolding an agent runs inside — so the agent finds and fixes its own mistakes before a human ever reviews. That is the mechanism that lets agents run longer without a developer watching every step, and the Radar also flags old disciplines like penetration testing as newly load-bearing because of AI. One caution from the same roundup of the research: agents overclaim finished work, and a review done only partway makes planted defects roughly twice as likely to slip through. The New Stack
“Security is going to be a big topic. Not even one year out.”
Agents and plans
Claude Fable 5.1 — Anthropic has released Claude Fable 5.1 — the same model as the tightly restricted Mythos 5.1, sold with standard safeguards. Where usage is billed by token it should cost an estimated 25% less than Fable 5 for typical workloads, and up to about 45% less for heavily agentic work, mostly through cheaper cache reads. Two things a buyer cares about: a real cut on the model your agents burn the most tokens on, and new Enterprise Frontier Safeguards that hold data in customer-controlled cloud — zero data retention in practice — while still screening for adversarial use. Anthropic
Claude usage credits — Owners of Team and seat-based Enterprise plans can now switch on usage credits, which let members keep using Claude, Claude Code and the Cowork assistant after their seat allowance runs out, billed at standard API rates. This fixes the mid-task wall — an agent stalling because someone's quota died — without upgrading seats. But it turns a fixed subscription into metered spend, so set the guardrails before enabling: limits can apply to the whole organisation, to a seat tier, or to one member. Team-plan credits expire 12 months after purchase and are non-refundable; members who hit their limit can request credits or a Premium seat from an admin. Claude Help Center
Village LLM Atlas — Village's LLM Atlas, which picks the strongest model at each of 201 price points, finds an open-weight model is the best buy at only 3% of budgets, down from 41% in June, after OpenAI's GPT-6 Luna and GPT-6.1 Terra took the cheap tiers back. If your routing plan assumed open models as the budget default, that window has shut for now: the closed labs are shipping cheap, strong small models faster than open labs can answer. Village
Cursor one-click deployment — Following this week's launch, the policy-relevant detail: Cursor publishes AI-built apps to live URLs on any paid plan with cloud agents enabled, after automated security checks whose ability to catch complex misconfigurations The New Stack questions — and individual subscribers can make their apps public. The precedent is cautionary: of 1,645 Lovable-built apps examined last May, 170 exposed sensitive information through misconfigured database permissions. Cursor provisions PostgreSQL databases automatically and says app secrets are kept away from the agent itself. The New Stack
Tools around the agents
GitLab 19.5 — GitLab 19.5, released October 15 and written up this week, makes its MCP server — the standard plug an agent uses to touch a platform — generally available, and lets Claude Code, Cursor and Codex register through the GitLab CLI so their sessions appear in the AI Governance dashboard and can stream to a SIEM (a security log store). If your team points agents at GitLab, you can finally answer which agent touched which project and when, from one place, instead of guessing. MCP calls on GitLab.com are rate-limited at 60 a minute for Free plans and 600 for Premium and Ultimate; a new auto mode that lets Duo agents run commands without per-command approval ships switched off by default. InfoQ
What people are saying
Does AI-fuelled bug-hunting make software better or just busier? IEEE Spectrum's account of PHP2Shell — an AI-assisted audit that turned a PHP crash bug, reported back in 2024 and ignored, into remote code execution affecting WordPress, Drupal, NASA and Wikipedia sites — pulled 1,257 points and 733 comments on Hacker News. The thread's split is the debate worth having with your leads: one camp argues that AI stress-testing means only provably correct code survives, and that only AI makes formal proof realistic at scale; the other says proofs still miss failures at the seams, and memory-safe languages plus honest testing do more than proof rituals. Repair is the bottleneck: four engineers spent two months writing the PHP patch after the exploits were disclosed. IEEE Spectrum
The big AI stories today
GPT-6 Cyber — OpenAI's gated GPT-6 Cyber found 81 unknown high-severity Firefox flaws and built 948 Linux kernel exploits — expect AI-found bugs in your queues to multiply. BleepingComputer
Microsoft buyouts — Microsoft is offering buyouts to 60,000 people as it shifts spending toward AI — watch it as the template for how big firms rebalance engineering headcount. New York Times via Techmeme
OpenSSH 10.6p2 — OpenSSH 10.6p2 fixes a password bypass on musl-Linux systems; the team says AI-generated bug reports now arrive in volume and will trigger faster releases. OpenSSH via Hachyderm
Google in Wyoming — Google pledged $17B for its first Wyoming data center, powered by 702 MW of gas fuel cells with carbon capture — growth continues despite voter backlash. Google
One takeaway
Set spend guardrails before the metered era lands. Two of today's threads point the same way: Anthropic's repriced IPO says vendors will keep monetising usage aggressively, and its new usage credits turn fixed Claude seats into metered spend. Decide this week who owns the agent budget and what each team's monthly cap is — the limits exist, but only if you set them. One thing not to worry about yet: cheap open models as your cost escape hatch; the best-buy data says that door is shut for now. Claude Help Center