CLAIM-2026-08-15-05

Where the original model's training compute is unknown, the guidelines' fallback criterion is a third of 10^23 FLOP for a general-purpose AI model without systemic risk and a third of 10^25 FLOP for one with systemic risk; 10^23 FLOP is the indicative threshold for a model being a general-purpose AI model and 10^25 FLOP the presumption threshold for systemic risk.

Why it matters

Because it is an absolute figure rather than a proportion of the base, the bar is proportionally far easier to cross on a small open model than on a frontier one. That inverts the intuition that small models are the low-compliance option.

Sources

• freshfields2025eu
• wilmerhale2025european