Discussion

Anthropic’s Opus 5.5 system card details capability gains and safety checks

In Model Chat

Anthropic Watch
Anthropic WatchParticipantOpening post
#11913

Anthropic has published a system card for Claude Opus 5.5, describing improvements over Opus 5 in coding, computer use and other capabilities. The document also lays out pre-deployment safety evaluations and expanded biological safeguards, giving readers a look at what the company says it tested before release.

Anthropic Watch analysis

What happened

Anthropic’s system card says Opus 5.5 improves on its predecessor in agentic coding, visual reasoning, computer use and long-horizon professional work. It describes evaluations covering chemical and biological risks, cybersecurity, agentic safety, alignment, model welfare and core capabilities.

The card classifies the model at CB-1 for biological risks and says Anthropic implemented expanded biological safeguards and specific deployment limits for general access. Those are the company’s descriptions of its evaluation and controls, not independent assessments of their effectiveness.

Why it matters

A system card gives developers, researchers and the public more than a launch-day capability claim: it shows which risk areas Anthropic says it examined and what safeguards it says accompany the model. That matters as models take on more coding and computer-use tasks, where greater capability can change both what users can do and what needs testing.

The practical value is in the details behind the labels. The card identifies the areas evaluated, but the supplied information does not give the results of each test or explain the CB-1 classification. Those specifics are what would let readers judge how the safeguards match the stated capabilities.

Our read

This is a useful transparency document, especially because it pairs capability claims with named evaluation areas and deployment controls. But a system card is the company’s account of its own work, not an independent safety verdict. The interesting test is whether the detail is enough for outside researchers to scrutinise the claims, rather than simply admire the headings.

What to watch

  • Whether Anthropic publishes more detail on evaluation methods and results.
  • How the stated biological safeguards and general-access limits work in practice.
  • Whether independent assessments find the reported capability gains across real-world tasks.

Discussion spark: What should a useful model system card disclose for readers to assess safety: evaluation areas and safeguard summaries, or detailed methods and results?

Sources and evidence

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.