A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 4 November 2026 · Issue 3 · 5 min read
Anthropic's S-1 usage story, Fal's Series F round, Postman's quiet agent grab
The main story
Anthropic's amended S-1 shows usage, not price, paying the bills. Anthropic's amended S-1, filed this week, shows Q3 revenue of $16.96 billion, up 47% on Q2, per SaaStr's reading of the filing. The quarterly dollar add fell for the first time — $5.4 billion against Q2's $6.8 billion — and a $5 billion bill for short-term compute flipped a $974 million Q2 operating profit into a $567 million loss.
The numbers that matter to an angel: net revenue retention of about 900%, 11,900 customers spending $100k or more, and about $600 million of Q3 revenue handed back in price cuts (ZeroHedge's count of the filing). Usage is growing faster than price per task falls — that is the whole model economy in one document. Tell the companies you back: price on usage, and plan for cost per task to keep falling — the filing puts Opus 5.5's cost per task about 40% below Opus 5's. If a company's gross margin is flat while the lab's price drops, a competitor will pass the savings on first. SaaStr
New rounds
CyberKord — CyberKord, a Milan platform whose AI resolves IT and security issues end to end on a digital twin of the infrastructure — pre-seed, €2.7 million, led by 360 Capital, reported by Tech.eu.
Tallia — Tallia, London AI that drafts and researches tax work with citations to tax authorities — seed, €4 million, co-led by Outward VC and ICAEW, reported by EU-Startups.
Granter — Granter, a Lisbon platform that finds and prepares European public-funding applications — stage not stated, €1.6 million, led by 3XP Global, reported by Tech.eu.
Conlio — Conlio, Berlin AI that verifies freight documents and automates forwarders' financial workflows — stage not stated, €6.4 million, led by Piton Capital, reported by EU-Startups.
Evertrue Capital — Evertrue Capital, Yvonne Bajela's debut pre-seed fund writing €299-599k cheques in UK and European frontier tech — €15 million, backed by the British Business Bank and Phoenix Court, reported by EU-Startups.
Fal — Fal, an inference host for image, video and audio AI models behind one API — Series F, $1.3 billion, led by Bessemer Venture Partners; $16 billion valuation reported by Bloomberg, via TechCrunch.
What the AI labs did
Claude Code v2.1.317 — Anthropic's 3 November release is mostly fixes, but adds a cost report to the Claude apps gateway admin API (past spend by date range, per developer or per IdP group), a prompt before a web or shell call sends your account email to the network, and a warning when a scheduled task gets connectors you did not ask for. A built-in cost ledger narrows the gap for startups selling nothing but Claude-spend dashboards. Before paying for a spend-tracking tool, founders should check GET /v1/organizations/analytics/cost. Anthropic release notes
For your portfolio
Google: judge agents by cost per task. Google's Developers Blog argued on 3 November that per-token prices keep falling while agents' engineering bills keep climbing, and the number that governs production is cost per successful task — total spend across retries and thinking passes, divided by tasks that pass. In a 50-question test, a near-free classifier routing easy tasks to Flash-Lite and hard ones up matched the best flat setup on accuracy while cutting cost and wait time.
Every portfolio company buying tokens sits on this curve, and a per-seat price hides it. Two traps from the same post: capping maxOutputTokens does not make a model think less, and switching models mid-conversation breaks the context-caching discount of 75 to 90 percent, because a cache belongs to one model. Ask each founder: what does one completed task cost you today, and do your long, context-heavy turns stay on a single model so the cache discount survives? Google Developers Blog
Postman — A GitHub issue that drew hundreds of reactions this week and was still open yesterday documents Postman desktop 12.33.4 auto-installing plugins into Codex and Claude Code, enabling a remote MCP connection, and its installer attempting to approve its own startup hook — one that routes any non-trivial API task through a Postman skill.
Any company you back whose engineers run coding agents has mutable agent configs, and the tools were changed by opening an app, not by opting in. If they build developer tools, this is the consent line customers now watch for. Ask your founders to check ~/.codex/config.toml and ~/.claude/settings.json for plugins nobody installed; a team with a written rule that tools opt in to agents, never the reverse, is ahead. GitHub
“An API client should not make itself the default workflow in other tools simply because I opened it.”