Five of the largest U.S. artificial-intelligence developers committed on Friday to a shared regime of pre-deployment safety evaluations, a joint incident registry and a binding disclosure timeline for frontier-model training runs, an industry-led effort designed to slow the state moratorium bills that have gained momentum since federal legislation collapsed last month.

The pact, announced in a joint statement issued in San Francisco and signed by OpenAI, Anthropic, Google DeepMind, Microsoft and Meta, would create a new nonprofit body, the Frontier Model Assurance Council, to administer common red-team protocols across signatories’ most capable systems. Nvidia and Amazon Web Services signed a separate cooperation memorandum pledging to align their hosting and chip-allocation policies with the council’s findings.

“We have heard the message from lawmakers and from the public, and we are answering it with verifiable commitments rather than reassurances,” OpenAI President Brad Lightcap said at a briefing held inside the company’s Mission Bay offices. “Independent evaluation is no longer optional in this industry.”

The announcement landed less than three weeks after the Sanders-Ocasio-Cortez moratorium bill died in the House Ways and Means Committee on a 24-21 vote, ending a three-month fight that briefly threatened to freeze new training runs above a defined compute threshold. Industry lobbyists celebrated the federal defeat at the time, but state legislatures in Sacramento, Albany and Olympia have since moved quickly to introduce parallel measures targeting data-center construction, training disclosures and grid use. A California Senate committee advanced a scaled-back version of that state’s bill last week.

Under the joint pact, signatories agreed to submit any model trained above 10^26 floating-point operations to a battery of dangerous-capabilities evaluations administered by the assurance council, with results shared confidentially with the U.S. AI Safety Institute and, in summary form, with state regulators that opt into a forthcoming information-sharing compact. Companies also agreed to a 60-day pre-deployment notification window, a marked shift from the closed development cycles that have characterized recent frontier releases.

A second component of the pact establishes a public incident registry to track real-world harms attributed to deployed systems, drawing on the model used by the Federal Aviation Administration’s safety reporting system. Reports would be filed by both companies and accredited third parties, and resolution timelines would be tracked publicly.

Anthropic Chief Executive Dario Amodei, who has long argued for mandatory pre-deployment evaluations, framed the agreement as a baseline rather than a ceiling. “This does not replace what legislatures choose to do. It gives them something to build on, and it gives the public something to see between now and then,” Amodei said in a statement.

Reaction from moratorium proponents was sharply mixed. Sen. Bernie Sanders, I-Vt., dismissed the announcement as “voluntary commitments from the same companies that spent forty million dollars to kill a binding law,” and reiterated his pledge to reintroduce a federal bill later this year. Rep. Alexandria Ocasio-Cortez, D-N.Y., struck a slightly different tone, calling the pact “an admission that the status quo was untenable” while warning that self-regulation had a poor track record across other industries.

State-level reaction was more pointed. California state Sen. Maria Elena Durazo, who is sponsoring the moratorium bill that cleared committee Wednesday, said the announcement would not slow her timetable. “We have heard variations of this commitment before, and we will continue legislating accordingly,” Durazo said by phone. New York Assemblywoman Latrice Walker, whose data-center disclosure bill is scheduled for a floor vote next week, said her office was “studying the framework but not pausing the calendar.”

Investors took the announcement in stride. Shares of the five signatories rose modestly in afternoon trading, with the Nasdaq composite closing up 0.6 percent. Analysts said markets had already priced in heightened compliance costs after the moratorium scare and viewed the joint structure as a manageable overhead.

“This is the industry choosing predictable rules over the risk of fifty different state regimes,” said Priya Ramaswamy, an analyst at Bernstein who covers hyperscaler infrastructure. “From a capex standpoint, that is a net positive, even if the headlines look defensive.”

The pact does not address the energy footprint of frontier training, an issue that has driven much of the recent state-level activity. Negotiations among the signatories collapsed late Thursday over a proposed cap on grid draw at data centers above 500 megawatts, according to two people familiar with the talks who were not authorized to speak publicly. Microsoft and Amazon Web Services pushed for the cap; Meta and Google opposed it, arguing that grid policy properly belonged to state public utility commissions and to the Federal Energy Regulatory Commission.

A senior White House official, asked about the announcement on the margins of an unrelated event, said the administration viewed the pact as “a useful complement” to the executive-branch framework President Donald Trump unveiled last month, while noting that the White House continued to view a federal preemption bill as the preferred long-term outcome. Trump’s framework, announced April 23, leans heavily on voluntary commitments and on the AI Safety Institute housed within the Commerce Department.

European regulators offered a cooler response. A spokesperson for the European Commission said the EU’s AI Office would “evaluate the substance of the announcement” but cautioned that voluntary U.S. arrangements could not substitute for the bloc’s binding AI Act obligations, which entered their next compliance phase in February.

The assurance council is expected to publish its founding charter and initial evaluation protocols within 30 days. Council leadership will be announced separately, the signatories said, with an interim director drawn from outside the five member companies. Industry officials said additional firms, including Cohere, Mistral and the Apple Intelligence team, had been briefed on the framework and were expected to weigh participation in the coming weeks.

Lightcap, asked whether the pact represented a final answer to the year’s regulatory turbulence, demurred. “It represents where we are today,” he said. “The work of building trust does not end with an announcement.”