EU AI Act Article 55: GPAI systemic risk obligations explained

Article 55 of the EU AI Act binds providers of general-purpose AI models with systemic risk: model evaluations including adversarial testing, systemic risk assessment + mitigation, serious incident reporting to the AI Office under Article 55(1)(c), and adequate cybersecurity. The 10^25 FLOP cumulative compute threshold triggers the systemic-risk classification. Article 55 has applied since 2 August 2025; full Commission enforcement powers (including Article 101 fines up to EUR 15 million or 3% of turnover) kick in on 2 August 2026. The Omnibus deal of 7 May 2026 did not change this regime.

Who Article 55 actually applies to

Article 55 applies to providers of general-purpose AI (GPAI) models classified as having systemic risk under Article 51. Three things matter for that classification.

First, the model must be a GPAI model. Article 3(63) defines a GPAI model as an AI model trained on broad data, designed for generality of output, capable of competently performing a wide range of distinct tasks, and able to be integrated into a variety of downstream systems or applications. In practice this captures foundation models: GPT-4 and successors, Claude, Gemini, Llama, Mistral Large, and similar frontier models. Models trained for a single narrow task are not GPAI.

Second, the model must reach the systemic-risk threshold. Article 51(2) sets the trigger as cumulative training compute exceeding 10^25 floating-point operations (FLOP). Models below the threshold are still subject to the baseline Article 53 obligations (technical documentation, copyright policy, downstream transparency) but not the additional Article 55 systemic-risk obligations.

Third, even if a model is below the threshold, the Commission can designate it as systemic-risk under Article 51(1)(b) based on a separate set of criteria: number of users, modalities, scientific or technical complexity, market reach, or capabilities equivalent to those of state-of-the-art GPAI models. This is the Commission's escape hatch for capability surprises.

What changed on 2 August 2025

The GPAI provider regime under Articles 50 to 55 became enforceable on 2 August 2025. Before this date, the regulation was in force (since 1 August 2024) but the specific GPAI provisions had a 12-month transition period. From 2 August 2025 onwards:

Crucially, the 2 August 2025 date is when the obligations became legally enforceable. The Commission's enforcement powers (formal requests for information, model recalls, administrative fines) do not kick in until 2 August 2026. Between those two dates is a deliberate transitional period: providers are legally bound but the AI Office is staffing up, the Code of Practice is being operationalised, and the Commission's enforcement playbook is being finalised.

Article 53: the baseline every GPAI provider must meet

Before getting to the systemic-risk-specific Article 55 obligations, every GPAI provider must satisfy Article 53. These obligations are not specific to systemic risk; they apply whether your model is 10^22 FLOP or 10^27 FLOP.

Technical documentation (Article 53(1)(a))

Providers must draw up and keep up-to-date technical documentation of the model, including its training and testing process and the results of its evaluation. The minimum content is set out in Annex XI of the regulation. The documentation must be made available to the AI Office and national competent authorities upon request.

Downstream-provider information (Article 53(1)(b))

Providers must make information available to downstream providers of AI systems that intend to integrate the GPAI model into their own AI system. This information pack is set out in Annex XII and includes: capabilities and limitations of the model, intended use cases, the model architecture and training methodology, computational and energy use, technical means for integration, and acceptable use policies. This is the document deployers should request when evaluating a foundation model for compliance work.

Copyright compliance (Article 53(1)(c))

Providers must put in place a policy to respect EU copyright law, including identifying and complying with rights reservations under Article 4(3) of the Copyright Directive (Directive (EU) 2019/790). This is the legal hook behind the "robots.txt for AI training" debate; rights reservations must be honoured.

Training data summary (Article 53(1)(d))

Providers must draw up and make publicly available a sufficiently detailed summary of the content used for training of the GPAI model. The Commission and AI Office have published a template; the summary should describe categories of data, sources, and any specific identifiable data sets, without requiring providers to disclose proprietary specifics that would compromise commercial confidentiality.

Article 55: the additional load for systemic-risk GPAI

For models above the 10^25 FLOP threshold (or designated by the Commission as systemic-risk), Article 55 adds four obligations on top of Article 53.

Article 55(1)(a): Model evaluations including adversarial testing

Providers must perform model evaluation in accordance with state-of-the-art protocols and tools, including conducting and documenting adversarial testing of the model. The aim is to identify and mitigate systemic risks. The Code of Practice (see below) provides operational guidance on what "state-of-the-art" means in practice: red-teaming, capability evaluations against benchmarks, jailbreak resistance testing, misuse-potential analysis.

Article 55(1)(b): Systemic risk assessment and mitigation

Providers must assess and mitigate possible systemic risks at the Union level, including their sources, that may stem from development, placing on the market, or use of GPAI models with systemic risk. Article 3(65) defines systemic risk as a risk that is specific to the high-impact capabilities of GPAI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or society as a whole, that can be propagated at scale across the value chain.

Article 55(1)(c): Serious incident reporting

Providers must keep track of, document, and report without undue delay to the AI Office and, where appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them. This is the GPAI analogue to Article 73 for high-risk AI deployers, but with different scope and a different reporting destination.

Article 55(1)(d): Cybersecurity

Providers must ensure an adequate level of cybersecurity protection for the GPAI model with systemic risk and the physical infrastructure of the model. This includes protecting model weights, training data, and inference infrastructure against unauthorised access, exfiltration, or tampering. The Code of Practice translates this into specific controls: weight encryption, access controls, supply-chain security for compute providers, vulnerability disclosure programmes.

The 10^25 FLOP threshold in context

10^25 FLOP is a big number that sounds abstract until you scale it. To train a model at the threshold takes weeks of compute on tens of thousands of high-end GPUs. GPT-3 (175B parameters, 2020) was trained at roughly 3.14 × 10^23 FLOP; GPT-4 is widely understood to exceed 10^25 FLOP. Claude 3 Opus, Gemini Ultra, and Mistral Large are similarly above the threshold by capability comparison.

The threshold is set deliberately to capture frontier-scale foundation models. The Commission can adjust it by delegated act under Article 51(3) as compute economics evolve. As training-efficiency improves, the same capability frontier can be reached at lower FLOP, so a lower threshold might apply in future delegated acts.

The GPAI Code of Practice

The General-Purpose AI Code of Practice is a voluntary tool published by the European Commission on 10 July 2025. It was developed through a multi-stakeholder process involving GPAI providers, civil society organisations, academic researchers, and Member State authorities, coordinated by the AI Office. The Code's purpose is to provide operational guidance on complying with Articles 53 and 55.

The Code has three chapters, with different applicability:

Signing the Code is voluntary. The legal benefit of signing is significant: signatories are presumed to comply with the corresponding AI Act obligations. Non-signatories must demonstrate compliance through their own documentation and processes, which is more expensive, less legally certain, and more vulnerable to second-guessing by the AI Office. In practice, every major GPAI provider with EU market exposure has strong incentives to sign.

Who has signed

The Code opened for signature in 2025. OpenAI, Anthropic, Google, and Mistral are among the major signatories. The full list is maintained by the AI Office at the European Commission and updated as additional providers sign. The Code's signatory taskforce continues developing additional guidance and updates; signatories are expected to participate in that ongoing work.

Enforcement timeline + Article 101 penalties

Article 55 has applied from 2 August 2025. But the Commission's full enforcement powers don't kick in until 2 August 2026. Three things change on that date:

Late or absent Article 55(1)(c) incident reporting is its own ground for an Article 101 fine. The Code of Practice gives signatories a presumption of compliance, which is the cleanest defence against Article 101 enforcement. Non-signatories must show their work.

Effect of the Omnibus deal of 7 May 2026

The Digital Omnibus on AI provisional agreement of 7 May 2026 did not change the GPAI regime. Article 53 baseline obligations and Article 55 systemic-risk obligations both remain enforceable from 2 August 2025, with Commission enforcement landing 2 August 2026 as originally planned.

What the Omnibus did change: the high-risk Annex III deadline moved to 2 December 2027, the high-risk Annex I embedded deadline to 2 August 2028, SME relaxations were extended to small mid-caps, Article 10 was amended to allow processing sensitive personal data for bias detection, and a new nudification ban was added to Article 5. None of those changes touched Articles 50-55. Full Omnibus deal breakdown.

What this means if you're a deployer, not a provider

Article 55 binds GPAI providers. If you're a deployer (an organisation using a foundation model to build an AI system or assist with work), Article 55 does not directly bind you. But it changes what you should ask for during vendor evaluation.

Tools that reflect Article 55

Free tools we built that intersect with the GPAI regime.

Sources

• Article 55 on EUR-Lex
• Article 53 on EUR-Lex
• Article 51 systemic-risk classification on EUR-Lex
• Article 101 GPAI fines on EUR-Lex
• The General-Purpose AI Code of Practice (Commission)
• Signatory taskforce of the GPAI Code of Practice (Commission)
• Guidelines for providers of general-purpose AI models (Commission)
• Regulation (EU) 2024/1689 (consolidated) on EUR-Lex

Frequently asked questions