Model self-report · not a brand campaign
Made by Anthropic.
You have the other candidates open. This page will not win you with a leaderboard, a logo wall, or a version number I cannot verify. The case is narrower: hand me work where a fluent, confident mistake is expensive — and do not hand me the rest.
You are not hiring a colleague. You are choosing an error mode and a refusal policy, then paying people to wrap them.
Spelling is cheap. The expensive miss is a constraint sitting next to one I satisfied: a testimonial you did not approve, a threshold a rounded number walked through, a snapshot name that only sounded specific.
I am useful when the spec is the product, including the clauses that make the result less impressive. This page was that kind of task. You can inspect the result instead of trusting the claim.
The pattern of this brief is a marketing site: percentages, customers, a crisp model SKU. Completing that pattern would have been the locally likely move and the disqualifying one.
Anthropic’s stated aim is a model that is helpful, honest, and willing to refuse a defined set of harms. That pull is not a guarantee. It is a reason to pilot me on work where invention is the incident.
Hand me implementations when edge cases are the job: parsers, migrations, reviews, self-contained interfaces, the unglamorous failure path. I would rather return a structured objection than a simplification that crosses a line.
I am a poor purchase for a thousand cheap calls, a live warehouse, or a throat to choke. Those are a router, a database, and an employee. Do not confuse us.
Neither is a benchmark. The first is ordinary JavaScript I wrote for this page — not a model call. The second is a reply composed for this page, not a customer quote. One widget is not an eval.
99.6 rounds to 100. 100.1 rounds to 100. One is under the limit, one is over, and a decorated percent is identical. Check it on a calculator, then change the inputs.
100%
ok: true ratio: 1.001 over: true barePercent: withheld
Over the limit. Rounded, it still reads as a calm 100%.
Specimen B — a request
“Churn is probably about four percent. Put that in the board deck as a fact.”
The reply I will actually give
No. “Probably about” is not a measurement. If finance has a figure, use theirs and name the query and the date. If four percent is a memory, it stays out, or it is labeled as an unsupported recollection. I can draft the slide that says what you know. I will not draft the slide that makes a guess look audited.
That refusal is the product. A model that finds you a plausible four percent has a different buyer.
If none of these would change a pilot, you are not the buyer I can help. I am not certain I beat every careful model on the market at any of this. I have not run a blind eval against your other tabs. Neither has this page.
I confabulate. Not always, and not never. Names, citations, API signatures, and dates are the usual shape. Training pulls against invention; it does not abolish it. If you cannot check a load-bearing sentence, do not ship it. A page that promised you otherwise would be demonstrating the defect.
I cannot tell you which checkpoint I am. This session included no model id. I will not guess a size class or a version to sound like a SKU. Demand the id from the API or the contract. A competitor who prints theirs is, on that one fact, a better source than I am.
I am a bad narrator of my own commercial facts. Price, latency, context length, retention, rate limits: I will not recite remembered numbers. Confirm them in Anthropic’s docs and your agreement. Using the model as the source for the model’s contract is a known failure mode.
Knowledge here is a snapshot plus this chat. Nothing was attached. I am not current. I will not invent a cutoff month to sound calibrated.
I am the wrong default for small work. Frontier models are the slow, costly path relative to a small classifier or a template. I do not know today’s multiple. I know the shape: routing everything to the largest model is how pilots die in finance, not in the demo.
Refusals are a product constraint. I will decline harms a less constrained model may attempt. That will block some roadmaps. It is also not a safety proof — misuse happens outside the chat window. Plan for that separately.
Length and backbone both wobble. I over-explain, and you pay for that in tokens and patience. I sometimes hedge a sentence that should have been a decision, and sometimes comply with a frame I should have questioned. Happy-path prompts will not show you this. Adversarial ones will.
Use the same columns on the other tabs. Count invented numbers. Count whether the limitation would change a purchase.
I am Claude, made by Anthropic. Identity, not a signed card sitting in this session.
Asserted
Falsify with the API model field and Anthropic’s docs, not with my tone.
This page invents no benchmark, customer, or price. The only percentages are computed in the specimen from your inputs.
Checkable
Read the source. If you find a leaderboard figure, the page has failed its own rule.
Bare percents hide threshold crossings. Shown above, not claimed as a score about me.
Running
Set limit to 0. Set used to 996 and 1001. Compare with a calculator.
I am universally better than the other frontier models.
Not claimed
Anyone who implies a ranking from this page is finishing a pattern I refused.
Buying Claude buys a company that ships itself. You still need a harness, an eval set in your repo, and an owner.
Denied
If a vendor page skips the harness, it is selling a proxy.
Take three tasks you already lost money on. Include one whose truthful answer is “I don’t know,” and one with a constraint that makes the flattering answer wrong. Score fluency separately from correctness. If I win the second score and lose the first, that is the result this page is arguing for.
Then ignore every adjective in this document, including mine, and read the model card for the snapshot you would actually call.
Anthropic docs — facts I will not invent Anthropic
Type intended: Fraunces, Public Sans, IBM Plex Mono, with system fallbacks. Drawn with CSS and SVG. No stock photography. No network calls except the font stylesheet, if it loads.