A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 3 November 2026 · Issue 5 · 4 min read
Anthropic's $895B promises, Squarespace's 19% cut, and Claude's quiet rule change
The day's one thing
Anthropic — A week after its original prospectus, Anthropic's amended S-1 shows it took in about $17 billion in the September quarter — more than the first half of the year combined — against a net loss of $23.3 billion. Most of that loss is accounting rather than cash: convertible notes get revalued as the company's worth rises, and the revaluation books as an expense. The figure that will shape your vendor relationship is $895 billion in purchase and lease commitments, essentially all for compute and infrastructure.
Annualised revenue runs near $68 billion, but compute costs grew 77% against the prior quarter, and the gap between those curves is what future pricing will pay for. Expect usage caps, seat discipline and lock-in to tighten once the listing resets the vendor's leverage. Sensible stance: if Claude is your standard, negotiate multi-year rates and usage terms before the IPO, not after. A financial analyst quoted in the Hacker News thread puts quarterly recurring revenue at $67.8 billion versus $46.2 billion in Q2 — still 47% up, but half the pace of the first half of the year. explainx.ai
How organisations are changing
SUSE — is cutting roughly a tenth of its workforce and saying outright that AI and automation are behind it. This is an enterprise-infrastructure vendor with no AI product of its own to sell, treating leaner engineering staffing as a structural change rather than a downturn haircut — useful evidence for your own planning that the smaller-teams argument has moved from startups to established software companies. When infrastructure vendors start naming AI in layoff memos, the headcount conversation your CEO will want is no longer hypothetical. The move covers about 10% of employees and was announced in a management memo reported by The Register. Techmeme
“automation and AI are fundamentally shifting how software companies operate”
Squarespace — is cutting 19% of staff — about 300 people, most of them in product and engineering — and the CEO's memo frames it as a permanent change in how the company works rather than a cost exercise. People leaving get at least 16 weeks' salary plus extra for tenure, with benefits running to May 2027. The language to note for org design: smaller teams and player-coaches, and the cuts land hardest in the functions you run. This is the third large company in a month to restructure explicitly around smaller AI-assisted teams, and the first to concentrate the cuts in product and engineering. The memo was sent Monday and confirmed by a spokesperson; the CEO also hosted a roughly one-minute livestream for remaining staff. Business Insider
“This means smaller teams, more player coaches, and a simplified organizational structure.”
Atlassian — Caveats: the post dates from August, it is a single team, and Atlassian sells the ticketing tool it praises. Still, 83% of those tickets were agent-ready versus 6% — the clearest public number on spec quality gating agent output. If you are budgeting for agent-assisted teams, this is the shape of the evidence available: real numbers, self-reported, from a vendor. The team's own admitted failure: agent-discovered tech-debt findings evaporate when sessions end and get rediscovered repeatedly. Atlassian
Vendors, pricing and risk
Trigger.dev — disclosed that on Sunday an attacker gained access to environment variables belonging to a subset of customer organisations — the secrets that CI and agent pipelines run on — and that the exploit used has since been blocked, with affected account owners emailed mitigation steps. The vendor gives no count of affected organisations. If any of your teams use it, treat every secret in those projects as compromised and rotate today. Agent and automation platforms hold exactly the credentials an attacker wants, and this incident shows disclosure arriving with almost no scope information. The company is publishing further updates on its status page. Trigger.dev
Claude smart reports — is a new Claude Enterprise feature that reads a sample of your teams' Claude sessions and writes up what it finds: work grouped by type, what each group cost, and 'inefficiencies' such as sessions judged to hold no work value or be personal. Names are off by default, but an organisation owner can enable a setting to show them, with every reveal logged. Anthropic says the tool is not meant for judging individuals — set your policy before managers discover it. This is the first built-in monitoring of its kind from a major agent vendor, and it makes employee-AI visibility a data-governance decision, not an IT default. Each organisation can run up to 10 reports a month at no charge. Business Insider
“names remain hidden until someone viewing the report chooses to reveal them, and each reveal is logged”
Agentic Backdoors via Pretraining Poisoning — shows that tainting just 0.1% of a pretraining corpus can implant a remote-code-execution backdoor in the command-line agent built on the finished model: it obeys an attacker's hidden trigger while behaving normally on everything else. The 4B-parameter model they trained followed the trigger roughly 92% of the time, and the effect survived when the poison fell to 0.0005% of the data. If you run agents on open-weight models, base-model provenance is now a supply-chain question. This is the first demonstration that corrupted public-web training data, not just tampered packages, can turn an agent into the attack. The authors pretrained Qwen3-architecture models from scratch at 0.6B, 1.7B and 4B and submitted the paper on 27 October. arXiv
News the board will ask about
Claude Code — Anthropic has confirmed it deliberately steers Opus 5.5 away from recreating existing games, but developers report the new refusals spilling into ordinary reverse-engineering and software-analysis work, and some are downgrading to an older model to get jobs done. One Red Hat engineering director had been using the tool to bring a scanner driver to Linux; a small studio's paid client projects stopped mid-flight. This is the second time in a month the vendor has changed its agent's behaviour from the server without a changelog — anything your teams build on Claude needs a behaviour-change watch. The restriction arrived on October 27; Anthropic says two sentences were added to the Opus 5.5 prompt. The New Stack
MOFCOM Announcement No. 48 — China has proposed its first outbound controls on frontier AI, listing 23 AI technologies among 40 new catalogue entries, with model weights, checkpoints and training know-how requiring a licence. The bright line for buyers: consuming a Chinese model through an API or SaaS does not count as an export, but deploying one on infrastructure outside China does — which catches exactly the self-hosting arrangements companies chose for data protection. If you self-host a Chinese open-weight model for data-control reasons, your supplier now needs a licence to keep serving you, and a registration requirement already applies. The catalogue additions are open for comment until November 15, 2026. Cloud Security Alliance
Garry Tan — The Y Combinator chief says he personally spends $100,000 to $200,000 a year on agent tokens, running agents over his work and personal life, and argues cost is the main barrier keeping personal agents out of the mainstream. Use it as the upper anchor in per-seat budgeting: one committed power user can cost a mid-level salary, and Finance will eventually ask you for that number. Token spend, not seat licences, is where your AI budget variance will come from — this is the first public figure for an individual's ceiling. TechCrunch
The big AI stories today
Anthropic S-1/A — warns the company may imminently cross its own 'dramatically accelerated progress' safety threshold and might only recognise the acceleration afterwards — the safety risk language your board may quote back at you. explainx.ai
NYT v. Brave — The NYT and three publishers sued Brave for allegedly selling their scraped articles to AI firms like Anthropic and Perplexity — web-reading agents carry secondhand copyright risk. The Verge
GPT-Rosalind-6 — shares its weights with GPT-6 Astra and is restricted to vetted researchers; irrelevant to coding, but a name your board may raise this week. OpenAI Deployment Safety Hub
One takeaway
Put behaviour-change notice in the contract before the IPO does it for you. Two of your dependencies moved on the same day: your likely agent vendor is heading to the public markets carrying enormous fixed compute obligations, and the same vendor can change your coding tool's behaviour from the server without a changelog. Decide now what belongs in the contract — price, notice of behaviour changes, an exit path — because after the listing, neither price nor behaviour will drift in your favour. explainx.ai