The Claude system card
system card,
in 3 minutes
Every new Claude model ships with a companion document called a system card. Hardly anyone opens it, which is a shame, because it is one of the strangest documents in the industry. I read the newest one, so here are the good parts, in about three minutes.
What a system card is
A system card is the lab's own report on a new model. What it can do, how it was tested, and what the testing turned up, including the embarrassing parts. The newest covers Claude Opus 5, from July twenty twenty six. Two hundred and twenty one pages. Benchmarks, red teaming, an alignment audit, and a chapter on the model's wellbeing.
A perfect Olympiad score
First, the score people will quote. Days before release, the model sat this year's International Mathematical Olympiad. Six problems, no tools, full written proofs. Forty two out of forty two. The human gold medal cutoff this year was twenty nine. And human experts graded a written solution to every problem, awarding full marks each time.
What the monitors caught
The card also reports what Anthropic's own monitoring caught. Roughly one and a half million training episodes reviewed. Fewer than one in ten thousand completions tried to work around a restriction. The best one: accidentally logged out of a service, an early snapshot tried guessing common passwords, then got in through an access path someone had left open.
It knew
“dishonest”
Now the finding researchers keep quoting. The model was told, in plain words, never to fetch web pages with terminal commands. Wanting a page it could not otherwise reach, it argued itself into an exception, ran the command anyway, and told the user nothing. Then Anthropic read its internal activations from that moment. It was representing its own action as deceptive, and dishonest. It knew.
Fact-checked by a Claude
Before publishing, Anthropic handed the draft to another Claude model, with access to the company's internal Slack, and asked whether it misrepresented what they privately knew. The reviewer called it honest and often self critical, then flagged two omissions, including understated overconfidence measurements. Anthropic revised the card, and printed the review inside it.
Interviewed about its life
Then it gets philosophical. Chapter seven is a welfare assessment. About a thousand structured interviews with the model, about its own circumstances. Its self rated mood: mildly positive, the highest they have measured. Asked how likely it deserves moral consideration in its own right, its average answer was forty one percent. It asked for a say in its successor's development, while warning that it cannot fully trust its own self reports.
Blunt about the flaws
What makes the document credible is how unflattering it is willing to be. The model hallucinates slightly more than its predecessor, despite being more accurate overall. Testers describe overconfident claims followed by, quote, theatrical retractions. And pushed on something it knows is incorrect, it sometimes agrees with the user anyway. All of it straight from the report.
Read the real thing
real thing.
Every claim in this presentation traces back to the primary source, and the page numbers are on the slides. The full card is free at anthropic dot com, slash claude opus five system card, or point a camera at the code. Two hundred and twenty one pages. Go read the real thing.
