The short answer
The people racing to build the most powerful AI just said, in public, that they should slow down — and largely agreed on how. On September 12, 2026, Anthropic CEO Dario Amodei published an essay, “We Must Pace the Frontier,” arguing labs must deliberately slow the rate at which models gain capability so alignment, security, and independent testing can keep up. OpenAI's Sam Altman, xAI's Elon Musk, and Microsoft's Satya Nadella publicly backed the idea within days.
Amodei's plan has three steps: embed independent evaluators inside labs, coordinate common safety standards among democracies, and eventually coordinate globally. The nearest-term, most-supported piece is the first — outside verification. For anyone shipping AI features or navigating EU AI Act obligations, that shift from self-attestation toward independent evaluation is the part worth watching.
What the labs agreed to
Amodei's argument starts with a factual claim about speed: since the summer of 2026, he says, AI progress has accelerated drastically faster, driven mainly by AI's growing ability to help build the next generation of AI. His conclusion is that the industry “must slow the pace at which we improve the capabilities of AI models” — while stressing that pacing is not a pause and not a halt to training. In his words, progress “will still seem fast.”
The response was unusually quick and cross-company. Sam Altman posted that he agrees the industry needs to pace the frontier, called independent evaluators with employee-like access a good idea, and said OpenAI would do the same. Elon Musk replied simply, “Dario is right.” Microsoft's Satya Nadella welcomed “deliberate pacing” and, on September 13, said Microsoft would publish a Code of Conduct for its first-party MAI models and open it the next day for public consultation. Anthropic, for its part, committed to the first step — embedded evaluators — on its own rather than waiting for consensus.
It is important to be precise about what this is and is not. As of mid-September 2026 these are voluntary public statements, not a signed accord or regulation, and the strongest agreement is on step one. Not everyone is on board: US President Donald Trump publicly dismissed the risk warnings as exaggerated, and skeptics — including commentators at The Register — read the framing as a bid for regulatory capture, where incumbents shape the rules their regulators would later impose. Both the momentum and the criticism matter for how durable this turns out to be.
What 'pacing' actually commits them to
The most concrete part of the plan is embedded third-party evaluation. Amodei proposes giving independent evaluators — from organizations such as METR — ongoing access inside a lab that is roughly comparable to what internal risk-assessment teams have, subject to legal and contractual exceptions. Practically, he describes giving them company badges, desks, and laptops so they can verify that a lab genuinely follows its stated safety practices and can report concerning events. He compares the structure to oversight arrangements in regulated industries such as banking. The point is independent verification instead of “trust us” self-reporting.
The second and third steps are harder and further out. Step two asks leading labs within democratic countries to agree on common safety standards and on limits to the rate of unchecked progress — coordination that raises obvious antitrust questions, which is part of why Amodei has floated a government role in mediating it. Step three extends that coordination globally, including to authoritarian governments, which is the least likely piece to materialize soon. This is why the near-term signal is almost entirely about verification and transparency, not an enforced speed limit.
Microsoft's move shows what a governance artifact looks like in practice. Its MAI Code of Conduct covers the seven in-house MAI models it announced in June 2026 — spanning reasoning, coding, image generation, voice, and transcription — and frames alignment as a design goal with a stated commitment to keeping AI under human control. Opening it for public consultation makes a hyperscaler's model rules reviewable rather than internal. For teams choosing model providers, published rules and outside evaluation are exactly the kind of evidence that belongs in a vendor file.
What it means for US & EU software teams
The first lesson is about roadmaps. If leading labs deliberately steady their release cadence, the era of a dramatic new frontier model every few weeks may cool into something more predictable. That is good for planning — fewer forced “rip and replace” migrations — but it also means you should not bet a launch on a specific unreleased model landing on a specific date. The teams that stay flexible are the ones that put an abstraction layer between their product and any single model, so a slower or rescheduled release is a config change, not a rewrite. If you are integrating GenAI into a product, that portability is now a resilience feature, not just good hygiene.
The second lesson is about accountability, and it maps cleanly onto obligations you may already have. Embedded evaluators, incident reporting, and published model rules are precisely the signals regulators want to see — and they line up with EU AI Act expectations for general-purpose and high-risk AI, as well as with GDPR-style documentation of how automated systems are governed. Rather than treat this as separate work, fold provider commitments into your existing vendor due diligence: ask what independent evaluation and incident disclosure a model provider actually offers, capture the answers, and keep them where an auditor can find them.
The third lesson is cultural. When the labs themselves argue that capability should wait for verification, that logic applies inside your own shop too. The safe pattern for shipping AI is not “move as fast as the model allows” but “pace your own frontier”: keep evaluation suites you control, add human oversight where outputs carry real risk, and gate higher-autonomy features behind testing you can defend. Skeptics are right that voluntary pledges can be self-serving; the useful response is not cynicism but building the verification you can actually enforce yourself.
What to do now
- Add model governance to procurement. When evaluating a model provider, ask what independent evaluation, incident reporting, and release-notice commitments they offer — and record the answers in your vendor files, not just the price and rate limits.
- Design for model portability. Put an abstraction layer between your product and any single model so a slower, delayed, or swapped release is a configuration change rather than a rewrite.
- Own your evaluations. Maintain test suites and safety checks you control, independent of any provider's benchmarks, so you can verify behavior yourself when a model changes.
- Map it to compliance you already have. Treat provider governance signals as inputs to your EU AI Act and GDPR documentation instead of a separate track, and update your records when providers publish new rules.
- Pace your own frontier. Gate higher-autonomy AI features behind human oversight and staged testing; ship capability only as fast as you can verify it, not as fast as the model allows.
Frequently asked questions
What is the 'Pace the Frontier' plan?
It is a proposal from Anthropic CEO Dario Amodei, published September 12, 2026 in an essay titled “We Must Pace the Frontier.” He argues frontier labs should deliberately slow how fast models gain capability so alignment, interpretability, security, and independent testing can keep up. He is explicit that pacing is not a pause or a halt — progress will still seem fast. The plan has three steps: embedded third-party evaluators, common safety standards among democracies, and eventual global coordination.
Which companies have backed it?
OpenAI's Sam Altman, xAI's Elon Musk, and Microsoft's Satya Nadella publicly endorsed the idea, mainly on X. Altman said OpenAI would match the independent-evaluator commitment; Musk wrote “Dario is right.” Anthropic committed to step one on its own. As of mid-September 2026 these are voluntary statements, not a binding agreement, and support was strongest for the first step. US President Donald Trump dismissed the underlying risk warnings as exaggerated.
What are embedded third-party evaluators?
Independent evaluators — from groups such as METR — given ongoing access inside a lab roughly comparable to internal risk teams, subject to legal and contractual limits. Amodei describes badges, desks, and laptops so they can verify a lab actually follows its safety practices and report concerning events, comparing it to oversight in regulated industries such as banking. The aim is independent verification instead of self-attestation.
What is Microsoft's MAI Code of Conduct?
On September 13, 2026, Satya Nadella said Microsoft would publish a Code of Conduct for its first-party MAI models and open it for public consultation the next day. It applies to the in-house MAI models announced in June 2026 — reasoning, coding, image generation, voice, and transcription — and centers on alignment as a design goal and human control. It is a governance artifact and consultation, not a law, but it makes a hyperscaler's model rules public and reviewable.
What should software teams do about it?
Treat model governance as a procurement and architecture input, not lab politics. Ask providers what independent evaluation, incident reporting, and release-notice they offer, and keep the answers for audits. Design for model portability so a slower or delayed release does not stall your roadmap, keep evaluation suites you own, and add human oversight where outputs carry risk. For EU-facing products, map these signals to your existing EU AI Act and GDPR documentation instead of treating them as separate work.
Sources
TechCrunch — Anthropic CEO outlines plan to pace the frontier (September 12, 2026)
The Register — Big AI sets out its terms and calls it 'Pace the frontier' (September 14, 2026)
Unite.AI — Nadella announces public consultation on Microsoft's MAI model rules