Editor's note. Today's thread is the price of intelligence. The lead is the clearest signal of the new mood: enterprises that spent two years told to use as much AI as possible are now demanding returns, capping budgets and switching to cheaper models. The second story is the supply side of that discipline — open-weight Chinese models that match the frontier at a fifth of the cost. The third is the one regulation main, and it earns the slot because today is the day: the EU Council is set to formally adopt the AI Act Omnibus, closing a thread this brief has tracked for weeks. Regulation otherwise stays out of the quick hits.
The bill arrives for "tokenmaxxing"
For two years the instruction inside engineering organisations was simple: use more AI. Employers incentivised developers to burn tokens without worrying much about the result, on the theory that capability bought speed and speed was worth almost any price. That era is ending. As OpenAI and Anthropic prepare for potentially historic IPOs, business leaders are no longer willing to spend without a clear picture of the return, and the shift from "tokenmaxxing" to efficiency is starting to bite the two labs that benefited most from the rush. Uber imposed AI spending tiers starting at $1,500 a month after blowing through its entire annual AI budget in four months; the CEO of agent startup Lindy moved 100% of its traffic off Claude to a cheaper alternative and expects to save millions. The analyst Gil Luria's blunt read is that current growth rates for both labs are the fastest they will ever see.
The reason this matters for a European enterprise is that cost discipline and sovereignty point the same direction. When the question stops being "what is the most capable model" and becomes "what is the cheapest model that clears the bar," the calculus that already favoured running your own infrastructure gets a second, financial argument behind it. A model whose per-token economics you control — and ideally whose weights you hold — is no longer the compliance compromise you accept reluctantly; it is also the line item your CFO is now asking you to find. The same pressure that is slowing the frontier labs' growth is the pressure that makes a model you operate yourself look less like caution and more like good budgeting.
Open weights answer the cost question
The cheaper alternative is increasingly an open one. Z.ai's GLM-5.2 — a 753-billion-parameter, MIT-licensed model with a one-million-token context window, released 17 June — lands within a point of Anthropic's Opus 4.8 on a key agentic benchmark while costing roughly $1.40 per million input tokens and $4.40 per million output, against $5/$30 for GPT-5.5 and $5/$25 for Claude Opus. On the Artificial Analysis Intelligence Index it sits at 51 — above Claude Opus 4.6, level with Gemini 3.5 Flash, and a handful of points behind GPT-5.5. The weights ship with no usage restrictions and no regional locks, which is precisely why they keep showing up in the migration stories behind the lead. This is framed as a cumulative position, not a fresh launch — but the position is now genuinely competitive at the top.
The strategic shape is hard to miss: the strongest US frontier model is rationed to roughly twenty government-vetted partners, while a frontier-tier open model is a free download. For an EU buyer the open route carries its own homework — Chinese-origin leaders bring provenance scrutiny, and a vendor's benchmark is not your production workload, so the discipline is to reproduce results on your own data before trusting them. But the structural point holds. A model you can pull down and serve on hardware you operate is the one option in this market that no trade policy, no waitlist and no budget freeze imposed by someone else can take away from you.
The Omnibus reaches the finish line
Today is the procedural endpoint of a long story. After the European Parliament approved the AI Act Digital Omnibus on 16 June, the Council is expected to formally adopt the text today, with publication in the Official Journal to follow — the last step before the redrawn timeline becomes law. The headline changes are the deferrals enterprises have been planning around: standalone high-risk systems under Annex III now face full compliance from 2 December 2027 rather than 2 August 2026, and the watermarking and transparency obligations for systems already on the market shift to 2 December 2026. The package also writes a new prohibition on AI-generated non-consensual intimate imagery directly into the Act.
It is worth being clear about what this is and is not. The Omnibus buys time on the highest-risk obligations; it does not repeal them, and the prohibited-practices and general-purpose-model duties already in force are unchanged. For a regulated buyer the practical takeaway is that the extra eighteen months are runway for getting the architecture right, not a reason to defer the decision — and the architectural choices that satisfy Annex III in December 2027 (documented data provenance, controllable retention, a clear line of jurisdiction) are the same ones that determine whether a deployment is defensible at all. The deadline moved; the work did not.
Quick Hits
- Mistral's sovereign Paris data centre comes online — the facility at Bruyères-le-Châtel, operated by French firm Eclairion, was expected to begin operations by the end of June with 13,800 Nvidia GB300 GPUs and 44MW of capacity. The chips are American but the location, operator and jurisdiction are European — the part of the stack Europe can build today. (Mistral is a European competitor; noted via third-party coverage.)
- GPT-5.6 stays behind the gate — OpenAI's flagship preview remains limited to roughly twenty government-vetted "trusted partners," with general availability still only "in coming weeks." OpenAI says it does not believe the government-access process should become the long-term default — a notable caveat from the company administering it.
- Gemini 3.5 Pro slips to July — Google has pushed general availability into next month, with early testers flagging token-efficiency and long-horizon issues; the 2M-token flagship stays in Vertex preview for now.
- OpenAI leans into advertising — at Cannes the company said roughly 20% of ChatGPT queries carry direct commercial intent, and ad-tech partners are lowering the barrier to entry — Criteo cut its campaign minimum to $10,000. Paid tiers stay ad-free, but the move is a reminder that on a hosted consumer model your queries are the product being monetised.
- MiniMax M3 keeps the open cohort deep — the open-weight M3, with a one-million-token context window, is part of a June wave that put a deep wave of frontier-grade open models into circulation. The frontier-tier open field is now several models deep, not a single standout.