The North Sea at night under a low sky, one distant light on the horizon

Field notes Governance

Your AI supplier may only see the jump after it has happened.

November 2026  ·  6 min read

On 2 November 2026 Anthropic filed an amended registration statement with the US Securities and Exchange Commission. The original filing of 26 October has already given us notes on insurance, the suppliers behind the supplier, agents that outrun their reviewers and regulators who can change the product. The amendment adds one sentence that deserves its own.

It says the company may soon meet a capability threshold in its own safety policy, and the paragraph around it says that kind of change has in the past been visible only afterwards. For a buyer, that changes what a supplier's safety policy is worth. A threshold you learn about after it has been crossed is an audit, not an alarm. Plan accordingly.

What the amendment says

The passage sits in a risk factor about the company using its own models to speed up its research, which it calls "closing the loop". It says its evidence on the pace of that acceleration "is derived from lagging indicators", and continues:

"As a result, we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred."

That sentence was already in the October filing. What the amendment adds comes next:

"For example, our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI.'"

Two qualifications follow, and they matter. The company says it has "not designated any of our current models as exceeding this threshold" and that it continues "to assess the underlying data". It also says it is "evaluating the implications of meeting the AI research and development capability threshold" and "will release further reports in the future" under the policy.

Elsewhere the amendment adds that the company does "not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding", and that it is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development". It is also considering policy changes "to require human review and approval of certain decisions" in its research as its models do more of that work.

What it does not say

Some of the commentary on this filing has run ahead of the text, so it is worth being exact.

It does not say any current model has crossed the threshold. It does not say what meeting the threshold will trigger: the company says it is still evaluating that. The human supervision line is a statement of what the company expects it can safely do, and the policy update it describes is under consideration, not adopted. Nor is this a promise to pause. The filing describes earlier precautions, including delaying pre-training of a new frontier base model "for more than a month" to strengthen security, and says future measures "could include additional measures that affect whether, how, and when we develop or deploy new models or capabilities". It does not commit to any particular one.

The threshold is also about the company's own research and development, not about the behaviour of the models you call through its API. Nothing in the filing says your service will change because of it.

Read it from the buyer's side

The useful line for a buyer is the one about precautions. The filing says the company "cannot predict with precision when we will determine that such measures are necessary, what measures will be required, or how long they will remain in place", and that as capabilities advance "such measures may become more frequent and more costly".

Put that next to the admission about lagging indicators. A supplier that measures its own progress after the fact will sometimes decide on precautions after the fact too. Release dates, availability and access terms can then move at short notice, for reasons that are about safety rather than about you. That is the policy working as designed. It is still a continuity event on your side.

The same point applies to every model supplier, including those that publish no thresholds at all. A supplier without a published threshold has nothing to disclose when it crosses one. Disclosure like this is only possible because a policy exists, which is a reason to ask every supplier for one, not a reason to move away from the one that wrote it down.

A supplier that tells you its measurements lag is giving you better information than one that says nothing. Use it: put the reading of that information on someone's job description.

What to ask your suppliers

Do you publish capability thresholds, and what happens when one is met? The filing says the policy "requires us to regularly publish detailed risk reports", and that the company's Frontier Safety Roadmap is "externally visible so customers and regulators can use it to gauge our progress". Ask each of your model suppliers for the equivalent. Where it exists, name an owner in your organisation who reads it when it changes.

What notice will you give when a precaution changes what we use? A safety-driven delay, a narrower release or a changed access condition reaches you the same way a regulatory one does. Ask what notice the contract gives for each, and get the answer in writing.

What do you keep human in your own development? The amendment says this supplier is considering requiring human review and approval of certain research decisions. Ask your suppliers what they already require, not what they are considering, and whether that is written into a policy you can read.

And on your own side: could you switch? If one supplier's precaution held back the model you depend on for a quarter, what would stop? Pin the model versions you run, keep an evaluation set you can run against a second model, and test the switch before you need it.

We are not neutral

ORCAHQ sells governance design, and we route work across several model suppliers, Anthropic among them. Read our emphasis with that in mind. Every quotation above is from the amended filing, and anyone can check it against the source.

We help organisations write the supplier questions into contracts and build deployments that can absorb a supplier's precaution without stopping. If a decision taken inside your supplier's lab could stop your operation, talk to us.

All field notes

Tell us what you are accountable for.

The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.

One reply from a person who has read what you sent.

Colophon

  • Plate 06, the sea at night: generated, Azure OpenAI gpt-image-1.5, 3 November 2026.

Every plate on this page is a generated stand-in for commissioned photography. The prompt that made each one is embedded in its file. How this site was made.