The safest way to judge an AI lab, or your own AI initiative, is not by what it says about itself. It is by what actually changed in what it ships and who gets to check it.
What happened in one nine-day window
On 12 September, Dario Amodei published an essay calling for what he termed pacing the frontier. He warned that AI agent swarms could threaten the stability of the internet within months, and he called for coordinated government action on chip restrictions and mandatory model testing. Sam Altman agreed publicly. He said OpenAI would give external evaluators employee-level access to its safety procedures. Elon Musk endorsed the essay outright. For a few days, three of the loudest voices in the industry were saying the same thing in public, slow down, and let outsiders check our work.
Ten days later, on 22 September at 9:38am Pacific time, Anthropic released Opus 5.5. The release notes described it, without apparent irony, as the company’s first model since it called for pacing the frontier. It was roughly 40 percent cheaper and 30 percent faster than its predecessor, and Anthropic pointed to a 680,000-line code migration completed in under a day as evidence of the jump. Eight hours later, at 6:02pm the same day, OpenAI released two new GPT-6 models, Sol and Luna, each more efficient than the version they replaced.
Was this hypocrisy?
No, and treating it that way misses the more useful lesson. The pacing pledge was never a promise to stop shipping. It was a promise to slow the size of capability jumps and to open up safety-verification access that did not exist before. Both labs kept shipping major upgrades within days of the pledge. That is not a broken promise. It is a gap between what a company says about its mood and what it actually does, and that gap exists in almost every organization, including yours.
The mood is not verifiable. The commitment is.
An essay is a mood. A blog post announcing restraint is a mood. A press release about responsible AI is a mood. None of it can be checked against anything, because there is nothing underneath it to check. What can be checked is whether an external evaluator actually received employee-level access to a safety review. What can be checked is whether a testing regime was documented, dated, and repeated. What can be checked is whether the thing that shipped eight hours after the pledge was reviewed under the new access or the old one.
Why leaders keep getting this wrong
Most executives read industry news the way most people read a company’s annual letter, as a signal of intent rather than a record of fact. Most people forming an opinion about AI have never properly used it, and the same pattern shows up at board level with public statements from AI labs. A CEO’s essay gets treated as a fact about the company, when it is really a fact about what the CEO wanted said that week.
This matters because boards make real decisions off this kind of signal. A board member reads a warning about AI risk and wants to pause investment. A procurement team reads a vendor’s safety pledge and treats it as due diligence already done. A leadership team reads that a lab has committed to caution and stops asking what actually changed in that lab’s process. In every case, the mood was treated as the commitment, and nobody checked whether the two matched.
What to check instead
When a vendor, a lab, or your own internal AI team makes a public claim about safety, restraint, or responsible practice, there is a short list of things worth actually verifying before you act on it.
Access. Did a named external party receive documented access to review the process, not just a statement that reviews happen.
Timing. Did the thing that shipped after the pledge go through whatever new process the pledge described, or the old one.
Change of practice. What concretely changed in how decisions get made, not how they get described.
Repetition. Is this a one-time statement or a standing, dated, repeatable process.
None of these require you to be a technologist. They require you to ask for the paper trail instead of the essay. The pace of the AI race is not going to slow because of a public statement, and treating a statement as if it were a guarantee leaves you exposed exactly when you thought you were being careful.
The same test applies inside your own company
This is not only about how you read OpenAI and Anthropic. It is about how you read your own AI initiative. If your organization has published an AI principles document, a responsible-use policy, or a governance charter, ask the same four questions of it that you would ask of a lab’s public pledge. Who has actually been given access to check the work. What shipped since the policy was signed, and was it reviewed under the new rules or the old habits. Which kind of AI leader you are right now is not decided by the policy document on the shelf. It is decided by what your teams were actually allowed to check before the last release went out.
The lesson from Opus 5.5 and GPT-6 arriving hours apart, both from labs that had just pledged restraint, is not that the pledges were worthless. It is that a pledge and a release are two different kinds of evidence, and only one of them can be audited. Judge vendors, and judge yourself, on the one that can.
If you are leading your organization through this, this is the kind of work I do with a limited number of senior leaders each quarter. → https://anglero.com/the-reset/
Strategic AI Advisor, MerkabaPhi AS, Oslo. 450+ keynotes across 30+ countries.
Listen to the Clarity at the Top podcast on Spotify.
Inquiries: anglero.com