Summary: Anthropic’s Responsible Scaling Policy 3.0, published February 24, 2026, removed the commitment to automatically pause development of a model deemed potentially dangerous. It instead pledged to match or surpass competitors’ safety efforts and separated Anthropic’s internal mitigations from recommendations for the field. A public Frontier Safety Roadmap and recurring risk reports increased disclosure without restoring an automatic stop.
Version 3.0 of Anthropic’s Responsible Scaling Policy landed on February 24, 2026, and the coverage followed at speed: TIME broke the story as an exclusive under the headline “Anthropic Drops Flagship Safety Pledge,” CNN cast it as the company ditching its core safety promise, and the Wall Street Journal ran it in print. A lab whose founding identity was safety-first had just overhauled the safety rule it was best known for.
Strip away the framing and a single sentence is what actually moved: Anthropic has ended its practice of automatically pausing development on a model deemed potentially dangerous. Under the new policy’s own logic, a developer that halts work to implement safety measures while rivals keep training and deploying without strong mitigations may be producing “a world that is less safe” — “the developers with the weakest protections would set the pace, and responsible developers would lose their ability to do safety research.” Per Anthropic, that language cleared CEO Dario Amodei and the board unanimously.
The Pledge Before This Revision
Back in September 2023, the RSP debuted as a voluntary if-then construct: cross a capability threshold (labeled in ASL levels) and the matching safeguards must already be in place. Its reach ran well past one company — OpenAI and Google DeepMind adopted comparable frameworks within months, and its imprint shows up in California’s SB 53, New York’s RAISE Act, and the EU AI Act Codes of Practice. Anthropic’s frontier models have run under ASL-3 safeguards since May 2025.
The clause carrying the weight committed the company to having stringent model security measures ready before training started — the guarantee TIME called “ironclad.” Not once did it actually trigger a pause. Yet the whole case for voluntary safety hung off that single clause.
The Substance of RSP 3.0
In practice, three things moved.
First, the automatic pause has been deleted outright. The new text says leadership would weigh delaying development only when it simultaneously believes Anthropic is the race leader and catastrophe risks are significant; what Anthropic pledges instead is to “match or surpass” the safety efforts of competitors.
Second, the policy now runs on two separate tracks: mitigations Anthropic applies to its own systems, and a distinct set it recommends to the field at large — so the company’s internal conditions stop being quietly exported as everyone else’s obligations.
Third, two new artifacts arrive: a Frontier Safety Roadmap and recurring Risk Reports. Public and nonbinding, the roadmap lists goals alongside graded progress — the announcement’s own examples include a moonshot R&D program for information security, automated red-teaming that surpasses the contributions of human bug bounties, measures for adherence to model constitutions, AI analysis of centralized records to catch insider threats, and a “regulatory ladder” policy roadmap. Risk Reports arrive every three to six months, each edition covering which model properties bear on catastrophic risk and where mitigations stand; in defined circumstances, outside expert reviewers are granted unredacted access.
Why Anthropic Says It Did This
Co-founder and chief science officer Jared Kaplan didn’t mince words with TIME: “We felt that it wouldn’t actually help anyone for us to stop training AI models,” and “we didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments… if competitors are blazing ahead.”
The announcement adds structural arguments on top: capability evaluations live in a “zone of ambiguity” where models near thresholds without definitively crossing them — a situation the old mechanism had no answer for; reaching RAND’s recommended SL5 security standard unilaterally is “currently not possible”; and the politics around regulation have moved. TIME fills in the commercial backdrop — Anthropic had just pulled in $30 billion in February at a reported $380 billion valuation, on annualized revenue that, per TIME, has grown roughly tenfold each year. Mashable, citing Axios, reports that the Department of Defense had pressed Anthropic to open military applications, holding restrictions out as the threat.
What Outsiders Made of It
Chris Painter, METR’s director of policy, told TIME that labs are operating in “triage mode” and that “society is not prepared.” What specifically worries him: dropping binary thresholds swaps clear tripwires for risk that piles up slowly — a “frog-boiling” effect. For advocacy groups such as the Transparency Coalition, the move supplied yet more proof that voluntary corporate pledges won’t hold up, and that safeguards belong in statute.
The honest counterpoint deserves logging too: transparency is moving in the opposite direction — upward. The revised RSP, the recurring risk reports, and January’s fully public, CC0-licensed Claude constitution all sit on the “publish it and let people check” line. What got dropped is the hard commitment to stop; what stays is the soft commitment to disclose.
Takeaways for Developers and Enterprises
Three consequences demand attention. On procurement and governance: teams that leaned on vendor self-restraint as a backstop just watched race dynamics formally override it — red lines now belong in your own contracts and evaluation pipelines. On policy: codifying RSP-style thinking into state law (SB 53 in California, the RAISE Act in New York) costs none of its technical reference value, but the argument that voluntary frameworks are sufficient has been undercut by the framework’s own inventor — fresh ammunition for the pro-legislation camp. On industry dynamics: whether OpenAI’s and Google DeepMind’s parallel frameworks undergo similar revisions is the signal to watch next. Should all three loosen together, “race to the bottom” graduates from slogan to description.