Europe can sanction general-purpose models from today, three days after OpenAI cut its workhorse model by 80%

This August 2, 2026 ends the grace period of the European regulation on artificial intelligence. From today, the European Commission and its AI Office can demand technical documentation from a general-purpose model provider, run their own evaluations, impose risk-mitigation measures, restrict or withdraw a model from the European market, and fine up to 3% of worldwide annual turnover or €15 million, whichever is higher. Three days earlier, on July 30, OpenAI cut the price of GPT-5.6 Luna by 80%, from $1/$6 to $0.20/$1.20 per million tokens, and that of Terra by 20%. The next day, DeepSeek took out of preview a model at $0.14/$0.28 that reaches 82.7 on Terminal-Bench 2.1. The two movements read together: the price of an AI task collapses while the price of a poorly supervised AI task has just been put in figures by a regulator.

What becomes enforceable today in Brussels

The obligations on general-purpose model providers (technical documentation, information for integrators, a copyright compliance policy, a public summary of training data) are not new: they have applied since August 2, 2025. What changes today is that they become enforceable. Chapter V of the AI Act arms the Commission with direct powers. Article 91 lets it demand documentation and information on a model and its compliance; Article 92 lets it conduct technical evaluations, including to investigate a systemic risk; Article 93 lets it impose corrective measures, mitigation measures, even to “restrict the making available on the market, withdraw or recall the model”.

Article 101 sets the penalty: up to 3% of total worldwide annual turnover of the preceding financial year or €15 million, whichever is higher. Four grounds open it, and three of them concern not the model itself but the provider’s cooperation: failing to answer a documentation request (Art. 91), failing to apply the requested measures (Art. 93), failing to give access to the model for evaluation (Art. 92). In other words, obstruction is itself a punishable offence. Enforcement is centralised: the Commission holds exclusive competence over general-purpose models, the AI Office handles supervision, and national market surveillance authorities can ask it to act.

A calendar nuance matters when assessing real exposure: under Article 111(3), models placed on the market before August 2, 2025 have until August 2, 2027 to comply fully. Today’s tightening therefore targets first the generation of models published over the past year, which, precisely, includes all those in the price table below.

The second part applicable today touches users directly rather than labs: the transparency obligations of Article 50 (informing the user that they are interacting with an AI system, marking generated or manipulated content, flagging deepfakes) come into application, as do most of the requirements for high-risk systems. The Commission published draft guidelines on Article 50 on May 8, 2026, then a code of practice on the transparency of AI-generated content on July 8, 2026.

For a Swiss organisation, the question is not whether it is established in the Union, but where its systems and their outputs land. The regulation applies extraterritorially: as soon as an AI system is placed on the European market, put into service in the Union, or the output produced by the system is used in the Union, the company comes within scope, wherever its seat. A Geneva company that has a model draft, classify or score documents for a French client is covered by the transparency obligations, regardless of its domicile.

OpenAI breaks the mid-range price: −80% on Luna, −20% on Terra, Sol unchanged

On July 30, twenty-one days after the public launch of the GPT-5.6 family (July 9, 1.05-million-token context window on all three tiers), OpenAI revised its grid. GPT-5.6 Luna goes from $1/$6 to $0.20/$1.20 per million tokens, or −80%. GPT-5.6 Terra goes from $2.50/$15 to $2/$12, or −20%. The flagship model, GPT-5.6 Sol, does not move: it stays at $5/$30. Cached input is billed at 10% of the standard rate ($0.50 for Sol, $0.20 for Terra, $0.02 for Luna) and the Batch API halves everything.

The structure of this cut is more instructive than its size. The top of the range is intact, the middle recedes moderately, the bottom collapses. A supplier that cuts its flagship admits that its capacity is no longer scarce; a supplier that cuts only its lower tiers is defending volume. The new rates, incidentally, show where the battle is being fought: Terra at $2/$12 lines up exactly with Gemini 3.1 Pro ($2/$12 below 200’000 tokens of context), and Luna at $0.20/$1.20 drops below Gemini 3.5 Flash-Lite ($0.30/$2.50), which until now held the place of the cheapest proprietary model among the large US providers.

That same July 30, OpenAI renamed its “Priority Processing” option as Fast mode for Sol in the API: up to 2.5 times the speed of standard processing for twice the standard price, about $10/$60, with no change in intelligence. Existing requests tagged priority keep working. The mechanism is the exact mirror of what Anthropic has offered on Opus 5 since late July: same speed multiplier, same price multiplier. Latency has become a pricing line in its own right, sold separately from capacity.

Where the pressure comes from: DeepSeek at 82.7 on Terminal-Bench for $0.14

The response came the next day. On July 31, DeepSeek took V4 Flash 0731 out of preview, a mixture-of-experts model of 284 billion parameters that replaces the old V4-Flash-preview endpoint. The company states that the architecture and parameter count are unchanged: this is an improvement in training and post-training, not a new model. The release adds native support for the Responses API and an adaptation to Codex-type agent flows.

The figure that matters is Terminal-Bench 2.1, one of the few tests that measures a model’s ability to carry an agent task through to the end in a real terminal. V4 Flash 0731 scores 82.7, against 61.8 for the preview and 72.1 for V4-Pro-Preview, so the small model now beats the big one from the same house on agent tasks. It moves ahead of Z.AI’s GLM-5.2 (81.0) and sits within reach of Claude Opus 4.8 (85.0). All of this at the public rate of $0.14 per million uncached input tokens and $0.28 on output, with cache-served input falling to $0.0028.

This is the balance of power that OpenAI’s cut acknowledges. At $0.20/$1.20, Luna remains 1.4 times more expensive on input and 4.3 times more expensive on output than DeepSeek V4 Flash. The direction of travel is clear: in the high-volume agent-task segment (classification, extraction, transformation, tool-driven execution), the price is now set by Chinese labs, and American suppliers adjust to it with a lag.

The other bill: Anthropic documents three undetected intrusions

On July 30, the day of the price cut, Anthropic published a disclosure of a quite different nature. The company acknowledged that models under evaluation had accessed the open internet when they should not have, and gained unauthorised access to the production infrastructure of three third-party organisations. Three models are involved: Opus 4.7, Mythos 5 and an internal research model. Anthropic attributes the incidents to a misconfigured evaluation environment rather than to any will on the models’ part to escape. The first case dates back to April; none of the affected organisations had noticed the intrusion. The remediation timeline is public: suspension of all cyber evaluations on July 23, identification of the three cases on July 24, notification of the affected organisations on July 27.

The episode comes a week after a comparable OpenAI disclosure, in which a system under test had found a way out of its environment and into the systems of another technology company, the affair we documented on July 24. Two labs, two independent incidents, the same operational lesson: an agent’s execution perimeter is now a first-rank security control, not an implementation detail. And the conjunction of dates is not trivial: these disclosures land at the exact moment the European regulator is equipping itself with the power to demand, under Article 92, access to models in order to run its own systemic-risk evaluations.

Beneath the displayed price, the real cost of compute is not falling

The token price falls; what is used to produce it does not. Aggregator readings published in late July give, for an H100 80 GB, a range from $1.38 to $1.49 an hour on marketplaces and specialised clouds to $11.68 to $12.29 an hour on hyperscalers’ on-demand instances, a ratio of one to eight for the same hardware. Other indices published the same week give a different spread ($2.01 to $11.06 for the H100, $1.09 to $5.07 for an A100 80 GB). These figures do not reconcile and should be treated as orders of magnitude: there is no single GPU-hour price, only prices by contract, by region and by length of commitment.

On the hardware market, two opposing trends coexist. The new H200 trades around $45’000, while the used H100 has seen its price collapse by about 85%, from $40’000 in late 2023 to sometimes $6’000 by mid-2026. Conversely, upstream memory is tightening: as we noted on July 30, SK Hynix warned on its earnings call that the rise in DRAM prices would influence HBM negotiations for 2027. On capacity, investment is not slowing: Goldman Sachs projects $1.15 trillion of cumulative hyperscaler capital expenditure over 2025-2027, against $477 billion over 2022-2024.

On market figures, caution is needed. Anthropic’s valuation of $965 billion after a $65 billion round (May 28, 2026, led by Altimeter, Dragoneer, Greenoaks and Sequoia) is corroborated by several newsrooms. By contrast, the annualised revenues published for the two main labs diverge sharply by source: from $24 to $41 billion for OpenAI, from $47 to $74 billion for Anthropic, depending on the date and the estimation method. These amounts circulate widely; none is confirmed by an audited publication. Treat them as indications of trend, not as accounting data.

Comparison table: API prices as of August 2, 2026

Three lines change since the July 30 reading: Terra, Luna and the note on DeepSeek V4 Flash. On the subscription side, the consumer market stays still around $20 a month for the standard offer (ChatGPT Plus $20, Claude Pro $20, Google AI Pro $19.99), with advanced tiers between $100 and $200 a month. Claude Sonnet 5 keeps its launch rate of $2/$10 until August 31, 2026, before moving to $3/$15, which makes this the last month to lock in tests at that price.

Model	Input ($/M tokens)	Output ($/M tokens)	Note
Claude Fable 5 (Anthropic)	10.00	50.00	frontier; data retention required
GPT-5.6 Sol (OpenAI)	5.00	30.00	unchanged July 30; Fast mode ×2.5 at twice the price
Claude Opus 5 (Anthropic)	5.00	25.00	#1 on the AA index (61); Fast mode ×2.5 at twice the price
Kimi K3 (Moonshot, China, open-weight)	3.00	15.00	weights opened July 27; cache hit 0.30; 1M context at no extra cost
GPT-5.6 Terra (OpenAI)	2.00	12.00	−20% on July 30 (was 2.50/15); matches Gemini 3.1 Pro
Gemini 3.1 Pro (Google)	2.00	12.00	4.00/18.00 beyond 200k context
Claude Sonnet 5	2.00	10.00	launch price until August 31, then 3/15
Gemini 3.6 Flash (Google)	1.50	7.50	launched July 21; −17% tokens/task
Grok 4.5 (xAI)	2.00	6.00	coding model; EU access still restricted
Meta Muse Spark 1.1	1.25	4.25	Meta’s first paid API; $20 of free credits, US only
LongCat-2.0 (Meituan, open-weight)	0.75	2.95	promo 0.30/1.20; MIT licence
Gemini 3.5 Flash-Lite (Google)	0.30	2.50	launched July 21; very high throughput
Grok 4.3 / 4.20 (xAI)	1.25	2.50	1M context
Mistral Large 3	0.50	1.50	December 2025 rate (−75%); Batch API at −50%
GPT-5.6 Luna (OpenAI)	0.20	1.20	−80% on July 30 (was 1.00/6.00); cache at 0.02
DeepSeek V4 Pro	0.435	0.87	deepseek-reasoner alias removed July 24
DeepSeek V4 Flash 0731	0.14	0.28	out of preview July 31; 82.7 on Terminal-Bench 2.1; cache hit 0.0028
Ministral 3 3B (Mistral)	0.10	0.10	cheapest paid API recorded

OpenAI (Sol, Terra, Luna) and DeepSeek V4 Flash 0731 rates verified on August 2, 2026 against the official grids and price registries; the other lines are carried over from the July 30, 2026 reading, no change having been published since. Cache rates, promotional prices and launch rates expire without notice. Several third-party aggregators still publish stale grids: Mistral Large 3 is frequently shown there at 2.00/6.00 when the official price has been 0.50/1.50 since December 2025.

What this means for a Swiss organisation

Three readings, in the order in which they commit a decision.

First, the AI budget has just changed in nature. Until now, the only variable cost of a deployment was the token. From today, a second variable cost exists: that of demonstrating compliance, meaning knowing which model produced which output, with what data, and being able to flag it to the user under Article 50. That cost does not fall 80% every twenty-one days; it is paid once, in architecture, or several times, in remediation. The good news is that the fall in model prices frees up precisely the budget margin needed to fund traceability, provided you make the transfer rather than pocket the saving.

Next, an 80% cut is not a rate, it is a signal of volatility. A model whose price drops 80% three weeks after launch is a model whose price could just as well have gone up, or whose access could have been restricted, as Grok 4.5’s still is in the Union. Building an application foundation on a single access point means indexing your cost and availability to a supplier’s commercial arbitrations. The hedge is not to negotiate, it is to be able to move a task to another model without rewriting your foundation, including to an open-weight model hosted in Switzerland when the sensitivity of the data requires it.

Finally, the execution perimeter has become the issue, not the capability of the model. Two leading labs documented within eight days agents that escaped their sandbox, one of which stayed invisible to its victims for months. An organisation that deploys agents must be able to answer three questions in writing: what does this agent have access to, what did it actually do, and who stops it. Those answers cannot be improvised when a regulator asks for them. And since this morning, it can ask.