The GPAI Code of Practice (Article 56): What It Is and Who It Binds
The Article 56 GPAI Code of Practice: a voluntary route to meet Articles 53 and 55 until standards exist, and distinct from Article 95 codes of conduct.
The General-Purpose AI Code of Practice is the central compliance tool for providers of GPAI models — the large models, including foundation models, regulated under Chapter V of Regulation (EU) 2024/1689. Drawn up under Article 56, the Code is a voluntary instrument that lets a provider demonstrate compliance with the binding Article 53 (all GPAI providers) and Article 55 (systemic-risk GPAI providers) obligations until harmonised standards exist. Signing it is not mandatory; but a provider that adheres to it can rely on it to demonstrate compliance and gets a clearer path with the AI Office.
This page explains what the Article 56 Code is, how it works, who has signed, and — the distinction practitioners most often get wrong — how it differs from the Article 95 codes of conduct, which are an entirely separate, voluntary instrument for non-high-risk systems. If you build on top of a GPAI model rather than provide one, start with GPAI obligations and provider vs deployer.
The one-paragraph answer
Under Article 56, the AI Office facilitated a Code of Practice for GPAI model providers. It is voluntary: a provider can use it as the primary means to show it meets Article 53 (technical documentation, downstream information, an EU copyright policy, a training-data summary) and, for systemic-risk models, Article 55 (model evaluation, systemic-risk mitigation, incident reporting, cybersecurity). The final Code was published on 10 July 2025, and major model providers signed. Adherence is a recognised way to demonstrate compliance with the relevant obligations. It is a bridge until harmonised standards are adopted — not a permanent regime, and not the same thing as an Article 95 code of conduct.
Who this binds — GPAI model providers only
The Code is for providers of general-purpose AI models — the model layer (Chapter V), supervised by the AI Office and the Commission. It is not for the SME deployers who build products on top of those models, and it is not for providers of ordinary high-risk systems.
If you fine-tune, integrate or deploy a third-party model (OpenAI, Mistral, Google, Meta) into your own product, you are generally a downstream provider or deployer of a system, classified by use under Articles 5 and 6 — the GPAI Code does not bind you. Your interest in it is indirect: the upstream model provider's Article 53 duty is to give you the information you need to meet your own obligations, and a signatory provider is more likely to supply that cleanly. See the general-purpose AI and foundation-model explainers for where you sit.
What the Code covers — Articles 53 and 55
The Code is organised around the two layers of GPAI obligation.
All GPAI providers — Article 53. Every GPAI model provider must: keep up-to-date technical documentation of the model; provide information and documentation to downstream providers integrating the model; put in place a policy to comply with EU copyright law; and publish a sufficiently detailed summary of the content used for training. The Code's transparency and copyright chapters give providers a concrete way to evidence these.
Systemic-risk GPAI providers — Article 55. A model is classified as systemic-risk where it has high-impact capabilities; Article 51(2) sets a presumption based on training compute greater than 10²⁵ FLOP. Those providers face heavier duties under Article 55: state-of-the-art model evaluation and adversarial testing, systemic-risk assessment and mitigation, serious-incident tracking and reporting, and adequate cybersecurity. The Code's safety-and-security chapter operationalises these. See systemic risk and the training-data summary.
| Layer | Who | Core duties | Code chapter |
|---|---|---|---|
| All GPAI | Every GPAI model provider | Tech docs, downstream info, copyright policy, training-data summary (Art 53) | Transparency; Copyright |
| Systemic-risk GPAI | High-impact models (Art 51(2), 10²⁵ FLOP) | Evaluation, risk mitigation, incident reporting, cybersecurity (Art 55) | Safety and Security |
How the Code actually works
Three features define its legal effect:
- It is voluntary, but it is the path of least resistance. A provider is free not to sign and to demonstrate compliance by other means. But until harmonised standards arrive, a signatory can rely on the Code to show conformity, and the AI Office treats adherence as evidence of compliance.
- It is a bridge to harmonised standards. Article 56 frames the Code as an interim instrument. Once harmonised standards are published and a provider conforms to them, the standards carry the presumption of conformity; the Code fills the gap until then.
- It does not enlarge the law. The binding obligations are in Articles 53 and 55. The Code is a way to meet them, not a new set of duties. A provider that meets Articles 53/55 without the Code is equally compliant — it just carries more of the evidentiary burden itself.
The GPAI obligations under Chapter V have applied since 2 August 2025 (models placed on the market before that date have until 2 August 2027 to comply — a separate transitional provision). Enforcement of GPAI-provider obligations sits with the Commission, whose enforcement powers, including fines on GPAI model providers under Article 101 (up to €15M or 3%), have applied since 2 August 2026 — not the national-authority Article 99 route.
Who has signed
The final Code was published on 10 July 2025, and the major model providers — across the US and EU labs that supply the foundation models most European businesses build on — became signatories, with some declining or signing only parts. The AI Office maintains the list of adherents. For a downstream business, the practical signal is simple: a signatory model provider has committed to supplying the Article 53 downstream information and training-data summary you may need, which makes your own compliance easier to evidence.
The distinction that trips people up: Article 56 vs Article 95
These two instruments share the word "voluntary" and almost nothing else. Do not conflate them.
- Article 56 — Code of Practice. For GPAI model providers. A means to demonstrate compliance with the binding Article 53/55 obligations until standards exist. Facilitated by the AI Office.
- Article 95 — Codes of Conduct. For providers and deployers of non-high-risk AI systems, to voluntarily apply some of the high-risk requirements (or other commitments — environmental sustainability, accessibility, stakeholder participation) where they are not legally required to. There is no underlying mandatory obligation; an Article 95 code is purely a good-practice uplift.
| Article 56 — Code of Practice | Article 95 — Codes of Conduct | |
|---|---|---|
| For whom | GPAI model providers (Chapter V) | Providers/deployers of non-high-risk systems |
| Relationship to mandatory law | A way to meet binding Art 53/55 duties | Voluntary uplift with no underlying duty |
| Purpose | Bridge until harmonised standards exist | Encourage good practice beyond the law |
| Facilitated by | The AI Office | The AI Office / national authorities |
The shared confusion is understandable, but the consequence of mixing them up is real: a GPAI model provider that thinks its Article 53 duties are "voluntary because there's a code" has misread the law. The duties are binding; only the Code is voluntary.
Worked example: where two companies stand
Skaldmodel OÜ — a fictional 60-person Estonian company (~€9M turnover) — and a downstream user illustrate the split.
- Skaldmodel trains and provides its own mid-size language model placed on the EU market. It is a GPAI model provider with Article 53 duties (technical documentation, downstream information, copyright policy, training-data summary). Its training compute is well below 10²⁵ FLOP, so it is not systemic-risk and Article 55 does not bite. It chooses to adhere to the Article 56 Code to evidence Article 53 cleanly and simplify its dealings with the AI Office — a voluntary step that lowers its evidentiary burden.
- A 30-person customer builds a chatbot on Skaldmodel's API. That customer is a downstream provider/deployer of a system, classified by use (here, limited-risk, with an Article 50(1) disclosure). The GPAI Code does not bind it — but it benefits, because Skaldmodel's Code adherence means the Article 53 information it needs is readily available. Separately, the customer could adopt an Article 95 code of conduct to apply some good-practice measures voluntarily, since its system is not high-risk.
The verdict: the Code binds the model provider's evidence strategy, not the downstream builder. On penalties, the downstream firm's system-level exposure as an SME is capped at the lower figure under Article 99(6); Skaldmodel's GPAI-provider fines run separately under Article 101, outside the Article 99 regime.
How Confir helps
Confir is built for the SME deployers and downstream providers who build on GPAI models — it classifies the system you ship by its use under Article 5 and Article 6, derives your role, and runs the assessment for any high-risk system, rather than the model-layer Chapter V duties that bind the upstream provider. Where you rely on a third-party model, Confir records which model you use and the provider information you hold (the Article 53 downstream material), so your own file is complete. The engine is deterministic and rule-based — the same intake yields the same finding, with the rule that fired shown. To place yourself, start with provider vs deployer and GPAI obligations.
Related guides
- GPAI obligations explained
- General-purpose AI compliance
- Article 53 — GPAI provider duties
- Systemic risk
- Training-data summary
- Provider vs deployer
- Foundation-model compliance
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →