The evidence ladder for AI capability claims

A model ships and the announcement post has a table. Anthropic’s page for Claude Fable 5.1 lists nine benchmark figures, among them 52.6% on Terminal-Bench-Science 0.1 against 24.7% for the previous version, and prices the model at $10 per million input tokens and $50 per million output tokens [4]. You read the table and then you have to decide something practical and slightly expensive: whether to move a month of client work onto a different tool. Nothing in that table tells you the answer, because none of it was measured on your work.

The useful question is not whether the numbers are true. Usually they are. The useful question is how much of each claim you are taking on the vendor’s word, because that varies enormously from one claim to the next and almost nobody sorts claims that way. This guide gives you a ladder for sorting them, and one piece of homework that will do more for your decision than every benchmark table you will read this year. It is written for a solo operator or a small team choosing between subscriptions. If you are procuring models for a regulated deployment and need a compliance file, you want the vendor’s documentation package and a lawyer, not this.

Rank a claim by how much of it you take on trust

There are roughly five rungs. At the bottom is a sentence in marketing copy: better at coding, more reliable, our most capable model. There is no evidence attached, so there is nothing to check. Next is a demo, which is a curated example someone chose because it worked. Next is a number the vendor produced by running a public benchmark against its own model, in a configuration the vendor selected and usually did not publish in full. Next is that same benchmark run by an organisation that does not sell the model. At the top is an artifact: something the vendor hands you that your own software checks, without the vendor’s cooperation, and that fails visibly and specifically if the claim is false.

Almost everything you read sits on the third rung. Anthropic’s system card for Fable 5.1 and Mythos 5.1 says so directly: “The majority of evaluations of Claude Fable 5.1 and Claude Mythos 5.1 were run in-house at Anthropic” [5]. That is not a confession, it is how the industry works, and a vendor writing it down is doing you a favour. Once you know which rung a claim is on, you know what to do with it. Third-rung numbers are useful for ruling models out, useless for ruling one in.

A Lean certificate is what the top rung actually looks like

On 1 August 2026 OpenAI published ten results across mathematics and theoretical computer science, credited to “an internal version of Astra, our next major model” [1]. The results themselves are not the part worth your attention. Alongside each one, OpenAI published a formal proof written in the Lean proof assistant and committed the set to a public repository [1][2].

Lean describes itself as “an open-source programming language and proof assistant that enables correct, maintainable, and formally verified code” [3]. The distinction that matters here is that a Lean proof is not a description of a proof. It is the proof, written in a form a piece of software checks step by step.

What that buys a reader is concrete. The repository ships under Apache-2.0, pins Lean 4.32.0, and depends on mathlib; the build instruction is to fetch the mathlib cache and run lake build All, or to build a single formalisation by module name [2]. If a step in the argument is wrong, the build fails on your machine, in your terminal, and no part of that process involves OpenAI. There is no press office in the loop and no permission to request. That is the whole point of the format.

Notice also what the same page does not let you check. OpenAI states that “The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates” [1]. That figure may well be exact, but it is a third-rung claim sitting inches away from a fifth-rung one. The proof compiles or it does not. The token bill is something you take on trust. Two sentences, two completely different standards of evidence, on the same page.

The same benchmark run twice gives two different numbers

The gap between a vendor’s number and an independent one is not usually dishonesty. It is settings. Epoch AI, which runs benchmarks itself rather than collecting vendor figures, publishes a worked example: for one model, the vendor claimed “an accuracy of 65% on GPQA Diamond”, while across Epoch’s own 16 runs the mean was “0.55 ± 0.03” [6]. Epoch’s explanation is flat and probably correct: “We believe that these different scores are due to differences in evaluation settings” [6].

Ten points on a headline benchmark, from settings. This is why Epoch runs most models multiple times on each benchmark, publishes its task definitions, and holds “consistent and well documented settings across models” [6]. A single number with no confidence interval and no published configuration is not comparable to anything, including the same vendor’s number from six months ago.

Benchmark versions move too, and the version is usually printed right there in the announcement if you look. Anthropic’s Fable 5.1 figures name Terminal-Bench 4.0, OSWorld 2.0, and CursorBench 3.2.0 [4]. A score against version 4.0 is not the same measurement as a score against an earlier one, so a chart that plots a vendor’s models over two years is comparing a moving target with itself.

A performance percentage is a claim a regulator can test

If you have ever assumed a company would not print a specific accuracy figure unless it could back it, there is a case that should adjust the prior. On 28 April 2025 the FTC announced a proposed order against Workado, which had advertised its AI Content Detector as 98% accurate at telling AI-written text from human-written text. Independent testing put the actual accuracy on general-purpose content at 53%, and the company had trained the model only on academic content rather than the range of material it claimed. The FTC’s Bureau director put it in one line: “Consumers trusted Workado’s AI Content Detector to help them decipher whether AI was behind a piece of writing, but the product did no better than a coin toss” [7].

The remedy is the interesting part for your purposes. The order bars Workado from making effectiveness claims without “competent and reliable evidence” supporting them at the time the claim is made, and requires the company to retain that evidence [7]. That is the standard the law applies to a marketing percentage, and it is worth internalising as a reader: the burden sits on whoever made the claim. When a vendor prints a number with no method attached, you have not been given evidence. You have been given a number.

Twenty of your own tasks beat every table on the page

None of the rungs above measure your work, so the decision comes down to a test only you can run. Save twenty to thirty real tasks you have already completed, with the outputs you actually shipped, and re-run them whenever you are weighing a switch.

Anthropic’s own guidance on building evaluations is the practical version of this and applies just as well to a folder and a spreadsheet as to a test harness. Be task-specific and mirror your real distribution of work, including the edge cases: irrelevant data, overly long inputs, poor or harmful user input, ambiguous cases [8]. Structure the questions so grading can be automated where possible, using exact match or string match, then an LLM with a written rubric, and reserve human grading for what genuinely needs it because it is slow and expensive [8]. The advice most people get backwards is the one Anthropic prints under the heading “Prioritize volume over quality”: “More questions with slightly lower signal automated grading is better than fewer questions with high-quality human hand-graded evals” [8].

For a two-person business that means twenty saved prompts with their known-good answers, graded roughly, run against both the tool you have and the tool you are being sold. It is dull, it takes an afternoon, and it is the only evidence in this entire process that was gathered on the work you are paid for.

The limits section is the most informative part of a system card

When a vendor does publish an evaluation document, read it backwards. Results sections are written to be quoted. Limits sections are written by people who had to look at the failures, and they tell you what the vendor itself does not believe its numbers prove.

Anthropic’s Fable 5.1 and Mythos 5.1 card is a fair example. It records an external assessment by METR concluding that Mythos 5.1 “is likely unable to fully and reliably automate R&D for frontier projects spanning multiple weeks” [5]. It also states, of one biological evaluation, that “we are increasingly concerned that evaluation results are not meaningful proxies for the likelihood of the designed plasmids being biologically viable” [5]. A company writing that about its own test is telling you the test is weaker than its score suggests, which is more information than any results table contains.

Use the presence of that material as a signal in itself. A vendor that publishes external evaluations, names the evaluators, and prints its own doubts is operating differently from one that publishes a bar chart. Neither is proof, but only one of them is behaving like it expects to be checked.

What still goes wrong

The top rung is almost never available to you. Formal verification works because mathematics can be written in a language a checker understands, and nothing you do reduces to that. There is no Lean certificate for “writes client emails that sound like me” or “gets the invoice reconciliation right”. For the overwhelming majority of claims you care about, the ladder tops out at the fourth rung, and the honest move is to stop looking for external proof and go run your own twenty tasks.

Even a checkable artifact proves less than it appears to. OpenAI is explicit that the division of labour was not total: “We helped prepare the manuscripts and formalize the proofs in Lean, and we take responsibility for their correctness, while the mathematical arguments themselves were generated by our system” [1]. The Lean build confirms the theorem is true. It does not confirm that the model produced the formalisation unaided, and no amount of compiling settles that question. A verified artifact bounds what you have to trust; it does not remove trust from the picture.

Your own evaluation has its own failure modes, and they are the boring ones. Twenty tasks is a small sample, so a one-task difference between two models is noise you should ignore. The set goes stale as your work changes, and you will not notice, because it keeps producing numbers. And you can overfit to it exactly as a lab can overfit to a public benchmark, by quietly picking tasks the tool you already like happens to be good at. The defence is to add tasks when a job goes badly in real life rather than when you are shopping, so the set grows from failures instead of from preferences.