Frontier Models and General-Purpose AI (GPAI)
Why this topic matters
Section titled “Why this topic matters”“Frontier model / GPAI” is the central controversy of post-2023 AI governance. Once a single model’s capabilities span nearly every downstream application, the traditional “regulate by scenario” approach breaks down, and direct obligations on the model layer itself become necessary. But how to define “frontier”, who should evaluate it, and whether self-regulation is sufficient — here the three jurisdictions give the most divergent answers.
Conceptual distinctions
Section titled “Conceptual distinctions”- Foundation model (Stanford CRFM, 2021): a large-scale, self-supervised, broadly adaptable model.
- Frontier model (popularised by Anthropic / OpenAI / Google DeepMind and the Frontier Model Forum in 2023): the foundation model closest to the frontier of capability — typically high-compute, high-parameter-count.
- GPAI (general-purpose AI model): the statutory term in the EU AI Act (art. 3(63)), slightly broader than “frontier”.
- Systemic-risk GPAI: an AI Act sub-tier, with a presumption threshold at 10²⁵ FLOP.
Snapshot of dedicated obligations
Section titled “Snapshot of dedicated obligations”| Dimension | China | United States | EU |
|---|---|---|---|
| Dedicated concept | None (subsumed under “generative AI services”) | No federal concept (EO 14110’s 10²⁶ FLOP is revoked); California SB 53 defines “frontier model” | GPAI + systemic-risk GPAI (two tiers) |
| Compute threshold | None | California SB 53: 10²⁶ FLOP | ≥ 10²⁵ FLOP (presumption) |
| Ex-ante gate | Algorithm filing (mandatory for public-facing services) | None federally; SB 53 requires transparency reports + critical safety incident reports | Presumption of conformity + CE marking |
| Technical standard | TC260-003-2024 | NIST AI RMF (voluntary) | CEN-CENELEC harmonised standards (in development) |
| Code of conduct | None | 2023 White House Voluntary Commitments | GPAI Code of Practice (finalised Jul 10, 2025) |
| Penalties | Cybersecurity Law art. 68 / criminal liability | California SB 53: $1M / violation | 3% of global annual turnover (GPAI) |
Scholarly debates
Section titled “Scholarly debates”Theoretical foundations of frontier-AI risk
Section titled “Theoretical foundations of frontier-AI risk”- Bostrom (2014), Superintelligence (Oxford) — the foundational text on superintelligence risk.
- Amodei, Olah, Steinhardt et al. (2016), “Concrete Problems in AI Safety”: the earliest systematic treatment of alignment, robustness, and scalable oversight.
- Russell (2019), Human Compatible: a value-alignment framework.
- Hendrycks et al. (2023), “An Overview of Catastrophic AI Risks”.
- Bengio, Hinton, Stuart Russell et al. (2023), “Managing AI Risks in an Era of Rapid Progress” (Science open letter).
- Anderljung, Barnhart et al. (2023), “Frontier AI Regulation: Managing Emerging Risks to Public Safety” — the framework paper for frontier-AI regulation, co-authored across Anthropic / OpenAI / DeepMind / GovAI.
Who is qualified to assess frontier capabilities?
Section titled “Who is qualified to assess frontier capabilities?”- Optimist / libertarian camp: Yann LeCun (Meta), Andrew Ng, François Chollet — argue that frontier risk is overstated and that regulation suppresses innovation.
- Cautious / safety camp: Yoshua Bengio, Geoffrey Hinton, Stuart Russell, Dan Hendrycks — argue that frontier risk is serious and requires government action.
- Governance-design camp: Markus Anderljung (GovAI), Jess Whittlestone (CLTR), Toby Shevlane (DeepMind) — research specific regulatory tools.
Academic commentary on the EU GPAI provisions
Section titled “Academic commentary on the EU GPAI provisions”- Engler (Brookings) series: ongoing coverage of GPAI implementation.
- Helberger & Diakopoulos (2023): critique of the drafting process for the AI Act’s GPAI chapter.
- Hacker, Engel, Mauer (2023), “Regulating ChatGPT and Other Large Generative AI Models” (FAccT 2023): one of the intellectual sources of the GPAI provisions.
- Almada (2025): criticises the AI Act’s presumption-of-conformity mechanism — “standardisation bodies are being asked to carry too much lawmaking weight.”
Chinese scholarship on GPAI governance
Section titled “Chinese scholarship on GPAI governance”- Xue Lan 薛澜 and Zhang Linghan 张凌寒: China has no “GPAI” concept; scenario-based regulation plus filing performs the same function as a substitute.
- Matt Sheehan, “Tracing the Roots of China’s AI Regulations” (Carnegie, 2024): TC260-003 functions as the counterpart to the EU GPAI CoP, but via technical standardisation rather than a code of conduct.
- Olivia’s thesis: the structural reason China’s “inclusive and prudent” approach omits a compute threshold — industrial structure + catch-up position.
The US “no dedicated frontier governance” debate
Section titled “The US “no dedicated frontier governance” debate”- Casey Newton / Platformer: the federal vacuum amounts to abandonment of the public interest.
- AEI / American Action Forum: federal deregulation is necessary for innovation.
- Bradford: the US’s “private-power” logic permits industry self-regulation to lead.
- Haugen, McLaughlin, Zuckerman et al.: frontier models should be treated as public infrastructure and assessed publicly.
Four core controversies
Section titled “Four core controversies”1. The rationality of the 10²⁵ / 10²⁶ FLOP thresholds
Section titled “1. The rationality of the 10²⁵ / 10²⁶ FLOP thresholds”- Weak scientific basis: the thresholds are “convenient orders of magnitude to regulate”, not “capability-transition points”.
- DeepSeek shock (Jan 2025): frontier-level capability with an order of magnitude less compute — casting doubt on the threshold design.
- Compute is bypassable: distributed training, distillation, and post-training tuning can all lift capability on modest compute.
2. Open-source vs. closed-source divergences
Section titled “2. Open-source vs. closed-source divergences”- AI Act art. 53(2): open-source GPAI is exempt from certain documentation obligations, but systemic-risk GPAI is not exempt even when open.
- In practice: Mistral’s flagship uses the MRL licence (not fully open source), deflecting full open-source obligations.
- DeepSeek’s all-out open release → once weights ship, there is no “pause” to invoke.
- Scholarly argument: Meta / Mistral argue open source is more transparent; Anthropic / OpenAI argue closed source is more controllable.
3. Industry self-regulation vs. hard-law constraints
Section titled “3. Industry self-regulation vs. hard-law constraints”- Frontier-lab self-regulation has loosened across the board in 2025–2026 (see company pages):
- Anthropic RSP v3 removed the pause commitment.
- OpenAI Preparedness v2 was simplified.
- Google DeepMind dropped its military-use prohibition in 2024.
- xAI rejected the self-regulation paradigm outright.
- Academic consensus: in the absence of hard law, the floor of self-regulation is set by the least-regulated player — and this is the central argument for the EU AI Act and California SB 53.
4. The methodology of capability evaluation
Section titled “4. The methodology of capability evaluation”- Each company’s red-teaming evaluation is not comparable — separate methodologies, proprietary benchmarks.
- MLCommons AILuminate / AI Luminate: attempts to build an industry benchmark.
- UK AISI / US AISI (reorganised as CAISI): government-run pre-deployment testing for frontier models.
- Academic critique (arxiv 2509.24394): the OpenAI Preparedness Framework “does not guarantee any concrete mitigation”.
Industry-practice lens
Section titled “Industry-practice lens”Safety-framework comparison across the main frontier labs
Section titled “Safety-framework comparison across the main frontier labs”| Company | Framework | Core method | 2025–2026 changes |
|---|---|---|---|
| Anthropic | RSP v3 (Feb 2026) | ASL tiers (modelled on BSL) | Revoked pause commitment; separates “unilateral” vs. “industry-wide” |
| OpenAI | Preparedness v2 (Apr 2025) | Threat categories × High / Critical | Simplified thresholds (dropped Low / Medium) |
| Google DeepMind | FSF v3 (Apr 2026) | Critical Capability Levels (CCL) | Expanded: new TCLs + manipulation CCL |
| Mistral | No independent framework | Open source + GPAI CoP signatory | — |
| xAI | Weak / absent | — | Publicly rejects the self-regulation paradigm |
| Meta | Frontier AI Framework | Similar methodology | Ongoing iteration |
Models plausibly above the compute thresholds (Apr 2026)
Section titled “Models plausibly above the compute thresholds (Apr 2026)”> 10²⁵ FLOP (EU GPAI systemic-risk threshold):
- Claude Opus 4.x (Anthropic)
- GPT-5 / 5.3 / 5.4 series (OpenAI)
- Gemini 3 Pro / Ultra (Google DeepMind)
- Grok 4+ (xAI)
- Llama 4 series (Meta)
- Mistral Large 2/3 (Mistral)
- Likely: Qwen 3.5 (Alibaba), ERNIE 5.0 (Baidu).
> 10²⁶ FLOP (California SB 53 frontier threshold):
- Claude Opus 4.x, GPT-5 series, Gemini Ultra, Grok 4+, Llama 4 Max.
Three typical cross-jurisdictional compliance strategies
Section titled “Three typical cross-jurisdictional compliance strategies”- “One document for three jurisdictions”: Anthropic / Google DeepMind map their RSP / FSF onto the EU GPAI CoP + California SB 53.
- “Jurisdiction-tiered documents”: OpenAI / Meta prepare separate compliance documents per jurisdiction.
- “Abandon the US and EU markets”: DeepSeek and most Chinese firms focus on the domestic market.
Structural events in Q1 2026
Section titled “Structural events in Q1 2026”- Trump EO 14365 (Dec 2025): attempts to preempt California SB 53 and the Colorado AI Act.
- Digital Omnibus Proposal (Nov 2025): proposes a 16-month delay of the AI Act high-risk provisions to Dec 2027.
- Anthropic RSP v3 (Feb 24, 2026): triggers the “loosening of frontier self-regulation” debate (TIME critique).
- Google DeepMind FSF v3 (Apr 17, 2026): against the tide, expands the framework.
Related rules
Section titled “Related rules”- EU AI Act arts. 51–56; Annexes XI / XII / XIII
- GPAI Code of Practice (Jul 2025)
- Digital Omnibus Proposal (Nov 2025)
- Generative AI Interim Measures
- TC260-003-2024
- AI Safety Governance Framework 1.0 / 2.0
- Anthropomorphic Interaction Services Measures