Claude Opus 5.5: What the Public Evidence Shows—and Doesn't

What is being assessed

Anthropic announced Claude Opus 5.5 on September 22, 2026 and published its own benchmark results and capability claims. Those results are useful evidence about performance on the named tasks, versions and settings. They are vendor-reported results, however, and should be read as such—not as a general guarantee about what the model will do in every workplace or open-ended situation.

This is an evidence synthesis based on public material. TST LAB did not run its own tests of Claude Opus 5.5. The assessment below compares the public claims with the scope and caveats of the available external evaluation; it is not a hands-on review.

What METR tested

METR describes a bounded evaluation of five software and research-oriented tasks, run through the model API over ten business days. The tasks included building a small language model, working with a codebase, training a program, making a game-playing bot, and analyzing a solar-technology question. METR also describes a questionnaire and interview with Anthropic and a separate internal report whose underlying evidence was not shared publicly.

That work asks a meaningful but specific question: how capable was the model on this selected set of tasks under the stated evaluation process? It is not a survey of every profession, a controlled measure of all production deployments, or proof that the model can autonomously conduct broad AI research. METR itself characterizes the result as incremental and discusses uncertainty around the interpretation.

Independence and editorial process

METR says the evaluation was conducted under an unpaid agreement. Its report also states that Anthropic had an opportunity to review and edit the final text, and that METR signed off on the published version. That disclosure matters. The report includes an external evaluation process, but its publication was not wholly insulated from the developer’s editorial input. Readers should consider both the stated methods and that review arrangement.

This does not by itself invalidate the reported results. It is relevant context for judging how independently the claims were tested, documented and presented. The public materials let readers inspect parts of the method and limitations; they do not provide every underlying artifact or reproduce all tests independently.

Benchmark performance is not broad capability

An evaluation score describes performance on a defined task set. Change the tasks, tools, scoring rules, time budget or human intervention and the result may change. Even strong performance on coding or research tasks does not automatically establish reliability in an unfamiliar workflow, safe operation, or consistent performance across users and environments.

The broader research literature on AI evaluation warns about narrow task selection, weak documentation, contamination, noisy measurement and the gap between a benchmark and the construct it is meant to represent. These are not reasons to ignore benchmarks. They are reasons to treat each score as an answer to a particular question rather than a crown awarded to a model for general intelligence.

What this evidence does not establish

The public material reviewed here does not establish that Opus 5.5 is best for every task, that it will match benchmark results in a particular user’s production setting, or that it can replace a team. It does not establish that every result has been independently reproduced, nor does a limited task suite settle questions about long-term reliability, cost in a specific deployment, or failure modes outside the evaluated tasks.

For a real decision, start with the work to be done. Define representative tasks, include ordinary and difficult cases, record acceptable error, test the exact model and configuration, and compare it with the current process. Public evaluations can help choose what to test. They cannot make that local decision on your behalf.

Assessment

The public evidence supports a measured conclusion: Anthropic reports a new model with results on named benchmarks, and METR reports a bounded evaluation across five tasks with disclosed uncertainty and an editorial review opportunity for Anthropic. This is informative evidence, not a universal capability certificate. TST LAB’s practical interpretation is that the next useful question is not “Which model wins?” but “What does this model reliably do on the work we actually need?”