A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 3 November 2026 · Issue 2 · 5 min read
OpenAI merges ChatGPT and Codex allowances from December; Creandum's $750M fund; Squarespace cuts 19%
The main story
OpenAI merges ChatGPT and Codex into one metered allowance. OpenAI told subscribers, in an email seen by BleepingComputer, that from December one token-metered allowance will cover both ChatGPT and Codex on Plus, Pro and Business plans, as the agent features of Work fold into the main chat app; quick Instant replies stay outside the meter, and heavier tasks burn more of the pool. No start day or token amounts are published yet. Separately, Work and Codex allowances on Pro 100, Pro 200 and Pro 500 were doubled on October 29, per OpenAI's release notes reported by BleepingComputer. The direction matters more than the detail: OpenAI is pricing subscriptions by compute, with agents and chat in one budget — the same economics any flat-priced agent product runs into. Tell a founder you back who sells agents on a flat plan: meter heavy users before December, and copy how OpenAI does it. BleepingComputer
New rounds
Prismus — a Swiss deeptech whose Lupa platform turns research and patent data into industrial leads — stage not stated, €163,000 from Venture Kick, reported by Tech.eu.
Sortrace — Stockholm AI that spots mis-sorted household waste from cameras on collection trucks — stage not stated, about $800,000 led by Almi Invest, reported by Tech.eu.
AvalonX — continuous AI penetration testing for small and midsize businesses, launched October 30 — stage not stated, up to $10 million from Ridown Capital, company release.
42AI — Israeli behavioural world model that predicts individual decisions, deployed at brokerages — seed, $62.5 million from Infinity Group, reported by TechCrunch.
Trajectory — lets lawyers and other specialists improve open models from task feedback — Series A, $45 million led by Sequoia at $300 million valuation, reported by TechCrunch.
Creandum — backer of Lovable and Black Forest Labs closed its largest fund, $750 million, for seed and Series A — a fund close, reported by Tech.eu.
What the AI labs did
Claude Code — The change landed server-side around October 27 with no changelog note, and users report lawful work being refused — one studio's commissioned game modules stopped mid-build. Startups building on Claude inherit the risk: the lab can change what their product does overnight, undocumented, mid-project. Ask any Anthropic-dependent founder what the fallback is when behaviour shifts — a second provider, a pinned model, or nothing. GitHub
PlanAssist — DeepMind's Gemini-built planning tool, which checks householder applications and drafts officers' reports, expands this month from three pilot councils to fifteen across England for a year-long beta. Pilots showed up to 58% less officer assessment time and over 70% of reports accepted without edits. A reference procurement now exists for decision-support AI in government, with humans keeping the final say. Show any govtech founder how the deal was built — DeepMind plus Google Cloud plus Faculty — and ask who their partner is. Google DeepMind
For your portfolio
Model spend is now a line next to payroll. Business Insider, reading Anthropic's IPO filing, reports that the firm's own Claude consumption — priced at what Anthropic charges customers — was worth more than its monthly payroll in April once stock awards are set aside. It's a paper figure, not cash out the door. Still, internal use grew more than 250-fold between January and September, and Claude wrote over 90% of the new code Anthropic shipped in September.
It's the sharpest public marker yet of the new cost structure: revenue growing far faster than headcount, and model tokens becoming a spend category on the scale of wages. Whatever a backed company's margins look like today, model spend is inside them. Ask each founder what the firm spends on model tokens per employee per month, and whether finance owns that number or it just lands on a credit card. Business Insider
Squarespace cuts a fifth of staff as prompt-built sites take share. Squarespace is laying off around 19% of its workforce — about 300 people — confirmed to TechCrunch on November 2. The backdrop is the prompt-built website: TechCrunch notes Wix bought Base44 for $80 million last year, and Lovable raised $400 million in August at a $13.3 billion valuation. Squarespace has its own AI tools, Blueprint (2024) and the forthcoming Forma, and cut anyway.
An incumbent couldn't hold its cost base even while shipping AI features, and the fight drags the floor of website pricing down with it. Any backed company whose product includes a site or storefront builder is now competing with a restructured incumbent and lavishly funded AI-natives at once. Ask the founder which of their customers would stay if a one-prompt builder matched the output — and what data, workflow or distribution the prompt can't reach. TechCrunch
A customer's agent now decides churn on outage day. Jason Lemkin writes on SaaStr that when a core vendor went down for a full day with no status page and no email, SaaStr's internal AI agent built a working replacement the same day and recommended dropping the vendor for good. They switched, at least partly, that same day. SaaStr runs three humans and more than twenty agents in production.
The things that protected B2B vendors — annual prepay, switching costs, human inertia — mean nothing to an agent. When the buyer is your customer's agent, an outage plus silence ends the relationship that day, not at renewal. Ask any B2B SaaS founder: who posts to the status page within the hour, and how does the API behave when a customer's agents hammer it? If nobody owns that, the churn story is already written. SaaStr
YC's chief spends six figures a year on agent tokens. Garry Tan says he spends roughly $100,000 to $200,000 a year on tokens for his AI agents — his own figure, given on Walter Isaacson's podcast The Builders, per TechCrunch. His agents work through his email, files and notes, and YC runs an internal agent, Stray Light, holding context on every founder it has ever backed. He thinks cost is the main thing keeping personal agents from the mainstream.
A live data point for anyone weighing consumer-agent startups: the heaviest user in the Valley already spends enterprise-seat money, and everyone below him is priced out until model costs fall. Unit economics, not capability, is the gate. Ask any consumer-agent founder what a power user burning four figures a month does to gross margin — and what happens to that margin when the labs cut prices. TechCrunch
dsh-superagent — DeepSeek's Harness team published an MIT-licensed prototype on October 31: a proactive coding agent that sits on one GitHub repository, takes work from approved people, runs each job in its own session and answers with pull requests. It's labelled a research prototype, not an official product, yet the repo already shows about 7.1k stars.
Free, open harnesses press directly on startups selling coding-agent workflow as SaaS — the defensible part is shifting from the model to the surrounding workflow, permissions and integrations. Ask any founder selling agent tooling why a customer keeps paying once the harness is free — and make sure the answer isn't the model inside it. GitHub