What Changed on 2 August 2025 for GPAI Providers

2 August 2025 is the general-purpose model milestone. Four chapters of the AI Act became applicable that day, and the one that changed behaviour is Chapter V — but two provisions that people assume arrived with it were expressly held back, including the Commission’s power to fine model providers.

What switched on

Article 113(b) of Regulation (EU) 2024/1689 made the following applicable from 2 August 2025:

• Chapter III Section 4 — notifying authorities and notified bodies (Articles 28 to 39). Designation machinery, ahead of the systems that will need it. See notified bodies.
• Chapter V — general-purpose AI models (Articles 51 to 56). Classification, provider obligations, authorised representatives, systemic risk obligations, and the codes of practice mechanism.
• Chapter VII — governance (Articles 64 to 70). The AI Office, the AI Board, the advisory forum, the scientific panel of independent experts, and the requirement for Member States to designate national competent authorities — which they had to do by this same date.
• Chapter XII — penalties (Articles 99 to 101), with the express exception of Article 101. Member States had to lay down their rules on penalties and notify the Commission by this date under Article 99(1).
• Article 78 — confidentiality, which governs how authorities handle what they obtain.

The Article 53 obligations

Article 53 places four duties on providers of general-purpose AI models placed on the Union market:

• Technical documentation of the model, kept up to date, containing at least the elements in Annex XI, and made available to the AI Office and national competent authorities on request.
• Information for downstream providers who intend to integrate the model into their own AI systems — documentation that enables them to understand the model’s capabilities and limitations and to comply with their own obligations, containing at least the elements in Annex XII. This is the provision that makes the model layer legible to the system layer above it.
• A policy to comply with Union copyright law, including identifying and respecting reservations of rights expressed under Article 4(3) of the Copyright in the Digital Single Market Directive — the text and data mining opt-out. See the copyright policy obligation.
• A sufficiently detailed public summary of training content, according to a template provided by the AI Office. The Commission published that template in July 2025. See the training content summary.

Article 53(2) narrows this for open models: the documentation duties in the first two bullets do not apply to providers of models released under a free and open-source licence that allows access, use, modification and distribution, and whose parameters — including weights, information on model architecture and information on model usage — are made publicly available. The copyright policy and the training content summary still apply, and the whole exemption falls away for models with systemic risk. The boundaries of that carve-out are contested in practice; see the open-source exemption.

The Commission also published a General-Purpose AI Code of Practice in July 2025, drawn up under Article 56 as the mechanism by which providers can demonstrate compliance pending harmonised standards. Adherence is voluntary; a provider that does not sign has to demonstrate compliance by other adequate means. The documents sit on the Commission’s AI Act pages.

Systemic risk, and the 10^25 threshold

Article 51 classifies a general-purpose model as having systemic risk if it has high-impact capabilities evaluated on the basis of appropriate technical tools and methodologies, or if the Commission so decides. It adds a presumption: a model is presumed to have high-impact capabilities when the cumulative amount of computation used for its training, measured in floating point operations, is greater than 10^25. Article 52 requires a provider to notify the Commission without delay, and in any event within two weeks, where the model meets that threshold — with the option of presenting arguments that the model nonetheless does not present systemic risk.

For a model in that class, Article 55 adds obligations: model evaluation in accordance with standardised protocols including adversarial testing; assessment and mitigation of systemic risks at Union level; tracking, documenting and reporting serious incidents and possible corrective measures to the AI Office and, as appropriate, national competent authorities without undue delay; and an adequate level of cybersecurity protection for the model and its physical infrastructure.

What did not switch on

Two omissions are load-bearing and both are easy to state.

Article 101 was expressly excepted. Article 113(b) applies Chapter XII from 2 August 2025 “with the exception of Article 101” — the Commission’s power to fine providers of general-purpose AI models up to €15 million or 3% of worldwide turnover. That power applies from 2 August 2026. So the Chapter V obligations became applicable a full year before the Commission’s dedicated fining power for breaching them. The AI Office’s other supervisory powers under Chapter IX — requesting information, conducting evaluations, requiring measures — are a separate matter.

The high-risk regime did not arrive. Nothing about Annex III classification, Articles 8 to 15, technical documentation under Article 11, conformity assessment, CE marking, EU database registration, the Article 50 transparency duties, post-market monitoring or Article 73 incident reporting became applicable on this date. Those were set for 2 August 2026, and most of them are no longer there either: Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July 2026, moved the stand-alone Annex III high-risk obligations to 2 December 2027 and the Annex I embedded case to 2 August 2028. The Article 50 transparency duties stayed on 2 August 2026. A vendor questionnaire sent in late 2025 demanding a CE mark for a high-risk system was asking for something the Regulation did not yet require — and the same questionnaire sent today is asking for something that is now more than a year further away.

Models already on the market

Article 111(3) contains the transition rule for incumbents: providers of general-purpose AI models that were placed on the market before 2 August 2025 must take the necessary steps to comply with the obligations in the Regulation by 2 August 2027. Models placed on the market on or after 2 August 2025 have no such runway — they had to comply at placement.

Two practical consequences follow. First, a model family with versions on both sides of that date can carry different compliance deadlines for different versions, which is a documentation problem as much as a legal one. Second, the transition applies to obligations, not to classification: a pre-2025 model that meets the Article 51 systemic risk threshold is a systemic risk model now, with the compliance deadline in 2027 rather than a different status. Nothing in Article 111(3) suspends the Article 5 prohibitions, which have applied to everyone since 2 February 2025.