Statements on pacing

On 12 September 2026, Anthropic CEO Dario Amodei published an essay, We Must Pace the Frontier, announced on X (@DarioAmodei). Amodei argued that frontier laboratories should moderate the pace of model capability growth so that safety techniques and oversight have time to keep up.

The Guardian reports that OpenAI's Sam Altman backed pacing and promised comparable evaluator access. xAI founder Elon Musk posted a three-word reply on X (read the post) quote-posting Amodei's announcement: “Dario is right”.

The three-tier proposal

Amodei pointed to two developments to explain his concern: the growing role of AI in building future models and recent multi-agent cybersecurity incidents during testing. He proposed three steps: first, laboratories grant embedded third-party evaluators ongoing access to inspect models and pipelines; second, labs in democratic nations coordinate with government support; and third, governments pursue broader international coordination.

Amodei noted that pacing means allowing adequate time for safeguards and evaluation rather than halting model training or technical work altogether.

What these sources document

Laboratory Public statement What these sources document
Anthropic Amodei stated Anthropic's commitment to embedded evaluators (12 September 2026). Anthropic reports an agreement with METR to investigate evaluation incidents. That announcement alone does not establish permanent oversight of training.
OpenAI See the reported response above. OpenAI reported a two-week RL pause in August. Separately, METR documented an incident investigation; this does not establish the scope of a future standing evaluator program.
xAI Musk posted “Dario is right” (12 September 2026). Musk's post indicates agreement with Amodei's essay, but this post does not specify company development policies or operational practices.

Editorial questions for evaluators

In our editorial view, these statements highlight practical questions for assessing safety commitments:

Forecasts are claims, not verified outcomes. Warnings about rapid agent escalation or timeline estimates reflect laboratory risk hypotheses rather than established empirical findings. Evaluators need inspectable evidence rather than forward-looking projections.

Endorsements do not establish operational rules. Agreeing with an essay on social media does not tell readers which written rules, shared measures, or verification procedures will apply.

Evaluator independence requires clear agreements. While voluntary transparency provides useful insight, meaningful evaluation depends on clear access terms and independent publication rights rather than informal statements.

Sources