A house newsletter · daily · researched and written by the Kurzia AI
Monday 2 November 2026 · Issue 1 · 5 min read
Anthropic files its S-1; Mistral open-sources Large 4; record VC dollars chasing fewer deals
The story that matters most
Anthropic S-1 — Anthropic filed a registration statement with the SEC on October 26 for an IPO of its Class A stock on Nasdaq under the ticker ANT; the number of shares and the price range are still undetermined.
Per SiliconANGLE, from the filing, first-half 2026 revenue reached $16.3 billion against an $80.7 billion net loss, and reports put the raise as high as $100 billion, which would top SpaceX's June debut. As the first pure-play AI lab to file in the US, its reception becomes the benchmark every AI mark on your books — and OpenAI's planned 2027 listing — gets judged against. The filing itself warns the compute bill could look irresponsibly large at first glance; the megaround window for every later-stage AI company partly rides on how the street prices that. Anthropic
Rounds and exits
42AI. Israeli stealth company whose behavioural world model predicts how individuals decide, live at brokerages; seed; $62.5 million, funded by its founders and Israel's Infinity Group over three years; already seven figures of revenue.
Dream. ex-Palantir founders building world models that design structurally sound buildings; seed; $17 million led by Andreessen Horowitz; deployed across 10 projects in Texas and California.
Clark. autonomous AI employees that do audit work for CPA firms; seed; $11 million led by Touring Capital, $16 million total raised; more than 35 firms already use it.
Tutor Intelligence. warehouse robots run first by remote teleoperators, whose deployment data then trains the foundation model; Series B led by a16z; amount not stated.
Blacksmith. CI infrastructure purpose-built to check code that agents write; Series C led by Sequoia; amount not stated; ARR up 17x in 2026.
Guidewire buys Hyperexponential. insurance software maker Guidewire is paying $1.7 billion in cash, around $2 billion with a revenue earnout, for the UK agentic-underwriting startup backed by a16z and Battery; expected to close next year, per TechCrunch.
Omen AI acquires Aquatic Labs. combining two real-time water-sensing plays for data-center cooling in a $16 million cash-and-stock deal, per TechCrunch; Omen raised a $31 million Series A in June.
OpenAI trims its round target. OpenAI has reportedly cut its new round from $30 billion to $20 billion ahead of a planned IPO, with $11 billion secured so far and a tepid response from some large investors; reported by the FT and the Wall Street Journal, via TechCrunch.
Creandum closes Fund VIII. the Lovable and Black Forest Labs backer closed $750 million for its largest fund, to back seed and Series A in Europe and increasingly the US, opening a new San Francisco office.
Record dollars, fewer deals. Crunchbase News: global venture funding reached nearly $680 billion through Q3 2026, a record pace, while North American deal counts fell 16% year over year to the lowest since 2023.
What the AI labs and big platforms did
Mistral Large 4 — is now fully open-weight: a 1-trillion-parameter multimodal model, about 49 billion parameters active per token and a one-million-token context window, downloadable for self-hosting. Its API runs $0.70 per million input and $2.10 per million output tokens, undercutting closed frontier pricing; the weights land squarely with sovereign-AI buyers and squeeze thin wrappers built on open models, while opening self-hosted coding-agent and security tooling. The licence requires a separate commercial deal from anyone reselling the model as a managed service. Mistral AI
Gemini 4 Carbon — is Google DeepMind's newest frontier model, aimed at long-horizon engineering and cyber defence; access starts with a subset of cyber defenders under the Fairwind Program before a wider release. It launches at $2 per million input and $10 per million output tokens, rising to $4 and $20 once the introductory period ends — bake that step into any Gemini-based margin forecast. The staged rollout also signals US pre-release testing now paces flagship launches. Google says its own teams have run the model for days straight to root-cause bugs across large codebases. Google DeepMind
Genie 4 — is DeepMind's real-time world model, generating interactive simulations with native audio that you steer in plain language; it is out as a research preview to trusted testers. When access to Genie 3's Project Genie widened in early 2026, Unity fell as much as 22.5% in a single session — game-prototyping and pre-visualisation tools face another step down, while robot- and agent-training environments are the category it opens. Sessions stay coherent for roughly two minutes, double Genie 3's limit, per The Verge. Google DeepMind
For your portfolio
Grok Bot — Users burned through their weekly limits over Halloween weekend, some jumping from half-used to 100% with days left on the clock. On November 1, Elon Musk confirmed on X that this was deliberate, not a glitch: serving the agent costs more than it earns. The product rides inside subscriptions such as SuperGrok at $30 a month with no standalone price, while background agent tasks consume far more compute than a chat reply.
The context matters: SpaceX told investors in a filing last month that AI is 52% of its income and its Super Intelligence division turned a $431 million profit, so this is about agent unit economics, not company survival. Startup Fortune reports roughly 800,000 paid weekly users. Ask every founder selling agents on flat seats: what does your heaviest decile of users cost to serve, and would usage metering save the margin or cost you the account? TechCrunch
How AI Applications Earn Their Premium — An essay by Tugce Erten on a16z's newsletter, dated November 2, works through the question every AI application now faces: when base models keep getting cheaper and more capable, what stops a customer replacing your product with a direct prompt and some engineering? Her case is that the durable premium rests on doing the whole job better than the buyer could in-house, on context they cannot cheaply assemble themselves, and on charging for completed work — a reviewed contract, a resolved ticket — rather than for tokens.
The stakes are visible in her evidence: Harvey won a blind evaluation against Microsoft Copilot and Repsol's own internal prototype across 50 lawyers, and in the essay's August 2026 survey of 100 enterprise LLM buyers, 94% expect LLM spending to rise, with 37% expecting it to at least double. Question for each AI application in the portfolio: if your biggest customer rebuilt your core workflow in-house this quarter, what would it cost them — and which parts would they still miss? a16z newsletter