Editor's note. Yesterday's thread was where the value gets captured; today's is who controls the inputs — the data, the model, and the substrate you run it on. The lead is the freshest of the three: Cloudflare has set a September deadline after which AI crawlers get blocked by default on ad-supported pages unless they pay, which quietly repositions the open web from a free training corpus into a licensed one. The second story is the model layer getting cheaper again — Claude Sonnet 5 lands close to the top model at less than half the price. The third is the deployment substrate: open-weight models just cleared the US government's highest compliance bar, which is exactly the template Europe has to reproduce on its own soil. Regulation stays in the quick hits, where the EU AI Act's general-purpose enforcement clock is now four weeks out.
Cloudflare turns the open web into a paywall for AI
Cloudflare said on 1 July that from 15 September its default settings will block "mixed-use" crawlers — bots that scrape for AI training and agents under cover of a search crawler — from any page that carries ads, unless the AI company separates those functions or pays for access. The change applies to all new sites, all new customers and every existing free customer, which is a large fraction of the web sitting behind one provider. To make the point concrete, Cloudflare published its 2025 crawl-to-referral ratios: Google sends one visitor back for every 14 pages it crawls, OpenAI one for every 1,700, and Anthropic one for every 73,000. The implicit "we crawl, you get traffic" bargain that built the web has, on those numbers, simply stopped functioning for AI.
The deeper signal for anyone building on AI is that the training corpus is being fenced and metered in real time, and the terms are being set by infrastructure companies and publishers rather than by the labs. That has a direct compliance edge in Europe, where the AI Act already requires general-purpose model providers to document their training data and respect content owners' rights reservations; a world where access is priced and logged makes provenance auditable in a way that scraped-by-default never was. For a regulated buyer the question migrates upstream from "which model" to "can this model's makers show a corpus they were licensed to use" — and increasingly, whether you can point to data you control rather than data someone quietly hoovered up before the gate came down.
Anthropic's Sonnet 5 pulls the price ladder down again
Anthropic launched Claude Sonnet 5 on 30 June as, in its own framing, a cheaper way to run agents — and the benchmarks make the pitch credible. Sonnet 5 posts 63.2% on SWE-bench Pro, 81.2% on OSWorld-Verified and 80.4% on Terminal-Bench 2.1 (up from 67.0% for Sonnet 4.6), carries a 1M-token context window, and exposes selectable reasoning effort from low through "extra high." On most published measures it lands within striking distance of Opus 4.8. The price is the story: introductory rates of $2 and $10 per million input and output tokens through 31 August, then $3 and $15 — against Opus 4.8 at $5 and $25.
The pattern is now familiar enough to plan around: capability climbs while the mid-tier price keeps falling, and the introductory discount runs on a clock that ends just as the EU's enforcement regime begins. For a European deployment that argues in one direction — the deciding variables are cost and jurisdiction, not raw quality, and a stack that can swap the model underneath it without re-plumbing captures each price cut instead of being locked to a single vendor's timetable. Sonnet 5's own two-tier pricing, cheap until 31 August and then not, is a small reminder that on a hosted model the terms are the vendor's to change.
Open weights clear the highest US compliance bar
The counter-move to renting the frontier has always been to run open weights on infrastructure you control, and that argument just got a concrete proof point — in the wrong jurisdiction for Europe. AWS confirmed that OpenAI's open-weight GPT OSS models and NVIDIA's Nemotron family, running on Amazon Bedrock in the GovCloud region, are now FedRAMP High and Department of Defense Impact Level 4/5 approved. In plain terms, open-weight models have cleared the most demanding accreditation the US government applies to sensitive and classified-adjacent workloads — the clearest evidence yet that "open weights" and "regulated-grade deployment" are no longer a contradiction.
That is genuinely good news for the open-model thesis and a template European buyers should read closely, because it also draws the map of what is missing on this side of the Atlantic: the same models, accredited to the same bar, but hosted under European jurisdiction and outside the reach of US disclosure law. The technology is portable; the certification and the sovereignty are not. A model an institution can download, inspect and serve is only half the answer — the other half is a substrate that has passed the local equivalent of FedRAMP High and sits where European law, not a foreign statute, decides who can compel access.
Quick Hits
- EU AI Act general-purpose enforcement is four weeks out — the Act's full enforcement regime for general-purpose models — Commission requests for information, model access and penalties of up to €15 million or 3% of global turnover — activates on 2 August; providers leaning on the voluntary Code of Practice to show transparency, copyright and safety compliance are nearly out of runway.
- OpenEuroLLM's delivery date is close — the 20-organisation consortium is due to ship its first open models, dataset and evaluation code by 31 July; a first checkpoint won't be a frontier model, but it is the checkable test of whether Europe's flagship open effort ships on time.
- Gemini 3.5 Pro is still in preview — Google's 2M-token flagship remains a Vertex preview after slipping its June target a second time; a confirmed GA and price would put the frontier price ladder — already being pulled down by Sonnet 5 — squarely in play.