Annex XIII: Criteria for Classification of GPAI Models with Systemic Risk

Annex XIII provides the assessment framework the AI Office uses to determine whether a GPAI model has high-impact capabilities and should be classified as systemic risk. The criteria include:

• Number of parameters of the model
• Quality and size of the dataset used for training
• Amount of computation used for training the model (measured in FLOPs) — including the 10^25 FLOPs presumption threshold
• Input and output modalities of the model (text, image, video, code, etc.)
• Benchmarks and evaluations of the model, including state-of-the-art performance
• Number of registered users or reach
• Any other indicator of high-impact capabilities

The 10^25 FLOPs threshold creates a rebuttable presumption: models above it are presumed systemic risk, but providers may argue otherwise. Models below it can still be designated if other criteria demonstrate equivalent capabilities.

The 10^25 floating-point operations threshold was calibrated to frontier models at the time of legislative negotiations (roughly GPT-4-class training compute). Key considerations:

• It is a rebuttable presumption, not a hard boundary
• The Commission can update the threshold via delegated act as training efficiency evolves
• Distillation, data quality improvements, and architecture advances may reduce the compute needed for equivalent capabilities—the threshold may under-capture risk over time
• The AI Office can designate models below the threshold based on qualitative criteria

Providers should track both their absolute FLOPs and benchmark performance to assess classification risk.

• Article 51 — Uses Annex XIII criteria to define systemic risk; paragraph (2) establishes the FLOPs presumption.
• Article 52 — Procedure for classification (notification, designation, rebuttal) based on Annex XIII assessment.
• Article 55 — Additional obligations triggered by systemic-risk classification.
• Annex XI Section 2 — Documentation requirements triggered by classification (evaluation strategies, red teaming, architecture).
• Article 97 — Delegated acts allowing the Commission to update Annex XIII criteria and thresholds.
• Article 90 — Scientific panel that may issue qualified alerts based on Annex XIII analysis.
• Article 113 — Application dates (Chapter V from 2 August 2025).

The recitals in the same consolidated AI Act on EUR-Lex contextualise the 10^25 FLOPs calibration, the rebuttable presumption design, and the Commission's power to evolve criteria. Use the official preamble on EUR-Lex—do not rely on unofficial recital lists without checking sequence and wording against the authentic text.