The one-screen version: The week's news all points at the same founder decision — where your model spend goes and who gets to learn from your content. On October 26, Anthropic filed its S-1 to list on Nasdaq as ANT, disclosing about $16.3B in revenue for the first half of 2026 at an ~85% gross margin, roughly 80% of it usage-based. On October 30, the four biggest labs agreed to honor one robots.txt token that tells all of them at once not to train on your site. On October 29, Mistral open-weighted a ~1-trillion-parameter model. And OpenAI's models landed in Google Cloud's Model Garden. Four stories, one throughline: the layer you build on is consolidating its economics in public — price accordingly.
1. Anthropic put its books in front of public markets#
The S-1 is the document this column has been waiting for. After August's reporting and October's investor-day numbers, the company your agents likely run on has now filed publicly: Class A shares, Nasdaq, ticker ANT, per CNBC.
The numbers that matter to you, not to a hedge fund:
- ~$16.3B revenue for the six months ended June 30, up from about $1.16B a year earlier — a roughly 14x jump — per Yahoo Finance.
- ~85% gross margin, with revenue roughly 80% usage-based API and 20% subscriptions (Cryptobriefing).
- An $80.7B net loss for the period — but about $72.4B of that is a non-cash charge, an accounting remeasurement, not cash burn. Q2 operating income was reportedly positive at ~$974M.
Ignore the scary loss headline; it's bookkeeping. The line that should change your pricing model is the 85% margin on 80%-usage-based revenue. That is the unit economics of the exact layer most of us resell, and a vendor that just showed public markets it keeps 85 cents on the dollar has no structural reason to keep selling tokens below cost. The filing also keeps control inside the house: it's a public-benefit corporation with a dual-class structure that leaves the co-founders with a voting majority — so product direction isn't about to be handed to a quarterly-earnings committee.
What you do Monday: compute your own product's gross margin with LLM spend as its own line item, budget next year at the standard API rate rather than today's promotional one, and keep the model layer swappable so a price move is a config change.
2. One robots.txt line now opts you out of all four labs#
On October 30, OpenAI, Google, Microsoft and Anthropic agreed to honor a single shared signal: a robots.txt group whose user-agent is the token __usage_ai-training. Per Unite.AI, it comes from an Internet-Draft, draft-kelley-aipref-training-preference-signal-00, and introduces no new directives — it reuses ordinary robots.txt syntax:
User-agent: __usage_ai-training
Disallow: /Three things founders need to get right. It is voluntary — a preference signal, not an enforced block. It governs training, not retrieval: it won't pull you out of answer-engine citations, and it doesn't undo anything already collected or trained on. And a plain User-agent: * rule does not count — the token must appear explicitly (FourWeekMBA). Rollout varies: Anthropic has begun, OpenAI says "coming months," Microsoft names Feb 1, 2027, Google says 2027.
We wrote the implementation up as a standalone how-to: how to add the opt-out token to your site, including what to keep alongside it.
3. Mistral open-weighted a trillion-parameter model#
On October 29, Mistral released open weights for Large 4 ("Le Chonk") — a roughly 1-trillion-parameter mixture-of-experts model (about 52B active), with long context and native multimodality, posted to Hugging Face under a modified-open license, per TechCrunch and Mistral's own post.
For most founders this is not "fire your API and self-host a 1T model" — the hardware bill for that is brutal. It's a negotiating and sovereignty fact: there is now a frontier-class open weight you could run in your own environment for a regulated or air-gapped workload, which quietly caps what a hosted vendor can charge for the same capability. If you've already read our open-weight-you-can-actually-run guide, this is the top of that ladder — impressive, and still mostly a cloud-GPU decision.
4. OpenAI's models show up inside Google Cloud#
Announced October 21: OpenAI's frontier models and Codex workflows are coming to Google Cloud's Model Garden, purchasable through Google Cloud Marketplace, with usage billable against existing Google Cloud commitments, per Devdiscourse. It's a private preview with GA "in the coming weeks."
The practical upside for a founder sitting on committed cloud spend: you may be able to point that commitment at OpenAI usage without signing a new contract or moving billing. The strategic read: the labs are deliberately multi-homing across every hyperscaler, which is good for your leverage and another reason to keep your integration cloud-agnostic.
The throughline across all four: the model layer is consolidating its economics in public. Price your product as if the promotional era is already over — because the S-1 says it is.
