Latest issueMonday 2 November 2026 · Issue 1 · 5 min read
Fewer seed cheques in a record funding year; Mistral open-weights a trillion-parameter model; six new rounds
The main story
Record dollars, fewer seed cheques. Crunchbase published figures on November 2 putting 2026 venture funding at nearly $680 billion through Q3 — a record pace — even as North American deal counts fell 16% year over year in Q3, the lowest quarterly total since 2023. The shortfall sits almost entirely in seed and early-stage rounds.
The record is a late-stage story: about a third of Q3's money went to just 27 companies raising billion-dollar-plus rounds, while cheque counts at your end of the market shrink. Tell founders not to count on a friendly Series A market; ask each one to tie the next raise to milestones now, not the calendar. Anthropic's possible $2 trillion IPO this month may lift sentiment, but the real signal is whether deal counts follow the dollars. Crunchbase News
“The real measure of a broader startup recovery we'll be watching: whether deal counts start to rise alongside the dollars.”
New rounds
Marble — Munich AI coworkers that learn processes by watching staff work; pre-seed; $3.75M; co-led by Heartcore and Seedcamp; reported by Tech.eu.
Talky Labs — Madrid accounting automation that learns each firm's unwritten rules; pre-seed; €1.5M, closed in September; no lead stated; reported by EU-Startups.
Pixel Systems — Stockholm tech that spots AI-faked images and documents through one API; seed; $3.1M; led by Unconventional Ventures; ~$21M pre-money, per ArcticStartup.
Roleo — Estonian personal AI career agent from Veriff's co-founder; stage not stated; €2.9M; led by Practica Capital with Inventure; reported by Tech.eu.
Clark — New York autonomous AI audit employees for CPA firms; seed; $11M; led by Touring Capital, total funding $16M; company release via PR Newswire.
Dream — AI world models that design whole buildings for the construction industry; seed; $17M; led by Andreessen Horowitz; reported by TechCrunch.
What the AI labs did
Mistral Large 4 (Le Chonk) — On October 29 Mistral published the weights of its 1-trillion-parameter flagship model, priced at $0.70 per million input and $2.10 per million output tokens on its own API. The licence is a modified MIT that forces anyone reselling the model as a hosted service to buy a separate commercial deal — bad for thin wrappers, good for European firms that must self-host for data residency. Founders reselling model access should read the licence terms before assuming open means free. Mistral AI
GPT-6.1 Terra — OpenAI's new mid-tier model went generally available in GitHub Copilot on October 26, priced at $1 per million input and $5 per million output tokens — roughly half the previous Terra, per Simon Willison. It is aimed at everyday agentic coding and computer use, so startups running high-volume agent loops get that work done at a much lower unit cost. Founders still running agents on older mid-tier models should rerun their per-task cost model this week. Simon Willison
Genie 4 — Google DeepMind announced its real-time world model on October 27 as a research preview for trusted testers: interactive scenes steered in natural language, with sound and dialogue generated alongside the visuals. It puts Google on the base layer of world generation, squeezing independent world-model startups in a category that is already consolidating. Sessions run about two minutes, per The Verge; founders in gaming and simulation tools should plan to build above this, not against it. Google DeepMind
For your portfolio
Grok Bot — On November 1 Elon Musk confirmed on X that Grok Bot's weekly limits were cut on purpose because serving the assistant cost too much relative to revenue; users had hit their caps days early over Halloween weekend. Startup Fortune reports the product reached about 800,000 paid weekly users within roughly two months of its August 11 launch.
It is the plainest public admission yet that background-running agents burn far more compute per user than a flat monthly price covers — the same unit-economics trap for any portfolio company shipping agents on a per-seat plan. Ask any founder with an agent on a flat plan what one heavy user costs to serve each month, and whether the price still clears if that usage triples. TechCrunch
“The cost of serving Grok Bot was unsustainably high relative to revenue”