AI Agents for Lawyers

Glossary

General-Purpose AI Model (GPAI) — Definition for Lawyers

A general-purpose AI model under the EU AI Act is a model with significant generality, such as GPT-6 or Claude Opus 5. Definition, tiers, signatories, enforcement.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. Model versus system: the distinction that decides who owes what
  2. Two tiers of GPAI obligation
  3. Who signed the Code of Practice, and what enforcement looks like since August 2026
  4. Why a deployer should care anyway

A general-purpose AI model, under Article 3(63) of the EU AI Act, is a model with significant generality that can perform a wide range of distinct tasks and be integrated into downstream systems. GPT-6 Astra, Claude Opus 5 and Gemini 3.1 Pro are GPAI models; ChatGPT, Copilot and Harvey are AI systems built on them. The Act's model duties fall on the provider, not on the lawyer using the product.

Model versus system: the distinction that decides who owes what

The AI Act regulates two things. An AI system is software that infers outputs from inputs for a purpose; a GPAI model is the trained component that many systems share. Anthropic's line-up as of September 2026 (Claude Fable 5.1, Opus 5, Sonnet 5) is a set of GPAI models; Claude.ai, Claude Cowork, Harvey, CoCounsel Legal and Lexis+ with Protégé are systems built on those and other models. Harvey runs a multi-model stack with Opus 5 live since 24 July 2026 and Fable 5.1 since 1 September 2026; Thomson Reuters rebuilt CoCounsel Legal on the Claude Agent SDK (GA 20 August 2026); LexisNexis draws on Anthropic, Google and OpenAI models. When a firm licenses one of these products it is a deployer of a system. The model provider is one or two contractual layers away, which is why the GPAI rules stay invisible to lawyers until something goes wrong.

Two tiers of GPAI obligation

TierTriggerProvider duties (summary)Applies since
All GPAI modelsMeets the Art. 3(63) definitionTechnical documentation for the AI Office; information for downstream system providers; a copyright compliance policy; a public summary of training content (Art. 53)2 Aug 2025 for new models; models already on the market have until 2 Aug 2027 under the Act's own transition rule (Art. 111)
GPAI with systemic riskPresumed above 10^25 FLOPs of cumulative training compute, or designated by the Commission (Art. 51)Everything above plus model evaluations, adversarial testing, systemic-risk mitigation, serious-incident reporting and cybersecurity (Art. 55)2 Aug 2025
Open-source carve-outFree and open licence, weights and architecture public, no systemic risk (Art. 53(2))Exempt from the documentation duties; copyright policy and training summary still required2 Aug 2025

The open-source row matters for firms going the Latham route. Latham & Watkins is fine-tuning Nvidia's open-weight Nemotron 3 models on its own H200 servers (September 2026); the base model's provider enjoys the lighter regime, and whether a downstream fine-tune creates a new model with its own provider turns on the scale of the modification, a point on which a firm should take specific advice. Harvey's Tenet, a post-trained open-weight legal model in research preview since 20 August 2026, raises the same question for a vendor. For the practical trade-offs see open-weight model and the open-weight versus closed comparison.

Who signed the Code of Practice, and what enforcement looks like since August 2026

The GPAI Code of Practice, published on 10 July 2025, is the voluntary route to presumed compliance. OpenAI, Google, Anthropic, Microsoft, Mistral, Amazon, IBM, Cohere and Aleph Alpha signed in full; xAI signed only the safety and security chapter; Meta refused, citing legal uncertainties. A non-signatory is not in breach, but it must demonstrate compliance by other means and forgoes the goodwill the AI Office extends to signatories.

Since 2 August 2026 the AI Office can demand documentation, evaluate models, order corrective measures and fine up to €15m or 3% of global turnover; prohibited practices carry €35m or 7%; supplying incorrect information €7.5m or 1%. It can also have a model withdrawn from the EU market. The Digital Omnibus, Regulation (EU) 2026/1744, deferred the high-risk system dates but left the GPAI regime untouched.

Why a deployer should care anyway

Three reasons, none of them theoretical. Continuity: a withdrawal order or a supplier's own decision can remove a model from your vendor's stack overnight, as US export controls did when they suspended Fable 5 and Mythos 5 access from 12 to 30 June 2026. Provenance: the Article 53 training-content summary and copyright policy are the only documents that tell you what a model learned from, which bears on confidentiality and on IP advice. Transparency: Article 50 duties on synthetic content run through the model provider's tooling, and Anthropic reportedly added a statistical watermark to Fable 5.1 outputs for that reason (secondary source).

Questions worth putting to any legal-AI vendor, in writing:

  1. Which GPAI models sit under the product, and did their providers sign the Code of Practice?
  2. Is any of them classified as carrying systemic risk, and what incident-reporting commitments flow down to customers?
  3. What happens to the product, and to our data, if a model is withdrawn from the EU market?
  4. Does the contract let the vendor swap models without notice?

The vendor due diligence guide turns these into a questionnaire, and Module 6 of the AI Lab for Lawyers course covers the AI Act's model and system layers side by side. For the whole statute as it applies to firms, start with the AI Act for lawyers.

Frequently asked questions

Is ChatGPT a general-purpose AI model under the AI Act?

No. ChatGPT is an AI system; the GPT-6 Astra or GPT-5.6 model underneath it is the general-purpose AI model. The distinction matters because model obligations under Articles 53 to 55 fall on OpenAI as model provider, while system-level duties, including transparency and any high-risk requirements, attach to the system and to its provider and deployers.

Do law firms have any obligations for GPAI models?

Not directly. GPAI obligations bind the model provider, so a firm using Claude, Gemini or GPT through a vendor owes nothing under Articles 53 to 55. The firm remains a deployer of whatever system it uses, with Article 4 literacy duties now and Article 26 duties for high-risk uses from December 2027. The practical exposure is vendor continuity if a model is withdrawn.

What is a GPAI model with systemic risk?

A general-purpose model with high-impact capabilities, presumed where cumulative training compute exceeds ten to the power of 25 floating-point operations, or designated by the Commission. Its provider must run model evaluations and adversarial testing, track and report serious incidents, and secure the model against theft. As of September 2026 the frontier models from OpenAI, Anthropic and Google are the obvious candidates.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion