Agentic AI Arrived This Month — and Europe's Enforcement Powers Are Already Live

Ten days, four releases, one direction of travel

Between 1 and 10 September 2026, four major labs pushed new frontier or near-frontier models into general availability. Anthropic went first on 1 September with Claude Fable 5.1 / Claude Mythos 5.1, priced at $10 per million input tokens and $50 per million output tokens — comfortably the most expensive line-up on the board. Google DeepMind followed a day later with Gemini 3.8 Flash at an introductory $0.25 / $1.50. OpenAI's GPT-6 Astra landed on 3 September, bundled into ChatGPT paid plans and ChatGPT Voice rather than sold per token. DeepSeek closed the run on 10 September with V4.1-Flash: $0.30 input / $1.20 output at peak, half that off-peak.

Read that list again and the shape of the market is obvious. Cheap, fast, tool-calling models at the bottom. Expensive reasoning models at the top. Almost nothing in the middle being sold as a plain question-and-answer chatbot. The OpenAI, Anthropic and Google DeepMind release notes for this cycle all describe scheduling, tool use and multi-step execution alongside the usual benchmark charts — the interface is no longer the product.

September's token prices, converted to euro

Nobody publishes a single comparable table across vendors, so here is one. Converted at an indicative rate of €1 = $1.08 (rates move; treat these as order-of-magnitude, not invoices).

The spread between the cheapest and most expensive output token is now roughly 83×. That is not a rounding difference — it is an architectural decision about what you route where. A support-triage agent that fires 40 tool calls per conversation does not belong on a €46-per-million output model, and a contract-review pipeline probably does.

Autonomy is the part that actually changed

Pricing is the easy story. The harder one is that the deployment pattern shifted underneath it. Enterprise buyers have largely stopped asking for text-in, text-out chat and started asking for multi-step planning with tool execution — the thing the vendors are now calling agentic workflows. That shift is measurable in places. Financial-services platform Lendi Guardian reported a 40% increase in engagement after moving from recommendation-based AI to autonomous execution. Whether you read that as validation or as a metric defined to flatter the feature is your call, but the direction is real.

Scale is real too. ChatGPT sits at roughly 900 million weekly active users, of whom about 50 million pay. When a system at that scale gains the ability to call APIs, write files and chain decisions, the failure mode stops being "wrong paragraph" and starts being "wrong action, executed".

It is worth remembering how recently the industry treated this as a gated exercise. OpenAI's earlier agentic rollout was restricted to around 200 controlled enterprise partners before broader availability, with the model classified under a high-risk safety tier. That staged approach — sandbox, then partners, then everyone — is now the template rather than the exception. It is also, notably, a lot easier to run when your user base is 200 companies than when it is 900 million people.

This is the tension Ahmed Gamal Elsharkawy put on the table at the BRICS Summit 2026: development speed cannot be the only objective, and safety standards, evaluation, governance and human oversight have to advance at the same rate. Dario Amodei, Sam Altman and Demis Hassabis have each, in their own framing, said something similar. The uncomfortable detail is that everyone saying it is also shipping.

Europe stopped asking nicely on 2 August

Here is where the European reader gets something the US reader does not. Since 2 August 2026, the voluntary era is over. Providers of general-purpose AI models face legally binding transparency requirements under Article 50 of the AI Act, and the European Commission and the European AI Office now hold direct enforcement powers — meaning supervisory authorities no longer have to wait for a national regulator to act on their behalf. Draft guidelines and self-regulatory codes of practice have been superseded as the operating framework.

In practice, for a European company deploying an agentic workflow, that means three things worth acting on now:

• Know what your model is doing under the hood. Transparency obligations flow down the chain. If you integrate a GPAI model into a product, you need documentation of its capabilities and limitations — from the provider, in writing.
• Keep an audit trail per agent action. The regulator's first question after an incident will be what the system decided and on what input. Logs are not optional plumbing anymore.
• Check data residency before you scale. Agentic models call tools, and those tool calls often cross borders. If your inference runs on US infrastructure, say so in your records.

The UK-based AI Act tracker at artificialintelligenceact.eu remains the most readable plain-English mapping of which obligation applies when, and the Commission's own AI Office pages carry the enforcement notices.

The European answer quietly changed too

Worth flagging, because it affects what you can actually buy. The strategy of funding large sovereign LLMs to compete head-on with US frontier labs at scale has largely been abandoned. What replaced it is more pragmatic: open-weight adoption, domain-specific platforms — legal, healthcare, public administration — and deployment inside European sovereign cloud infrastructure.

Mistral AI remains the clearest European comparator. Mistral Medium 3.5, from 29 April 2026, sits at $1.50 input / $7.50 output — cheaper than Grok 4.6 on input, more expensive on output, and roughly a tenth of Claude Fable 5.1's output cost. If your workload is document-heavy and latency-tolerant, that is a defensible position. We track how European-available models and local options behave under load on our benchmarks at AI Arena, where the RTX 5060 Ti 16 GB rig handles the open-weight side of the comparison.

What a developer should do with this

Route by cost tier, not by vendor loyalty. Put cheap Flash-class models on the high-volume, low-stakes steps and reserve the €46-per-million output models for the decisions that justify it. Version-pin everything, because September 2026 alone changed the pricing on half your stack. Keep your own evaluation set — it is the only safety mechanism you control, and the only one that will hold up in an enforcement conversation. And if you are selling into the EU, treat the transparency documentation as part of your release checklist, not a legal afterthought.

The speed-versus-safety debate has been running for three years and produced a lot of conference panels. What changed this month is that one jurisdiction now has binding rules and enforcement muscle, and the models getting shipped into it are the ones that act without asking. That is a more interesting test than any benchmark chart.