A game can be only 26 KB and still pass a demanding little test: press R, watch the game retrace its prior history, and do that successfully three times out of three. That is what happened in our Claude Opus 5 Rewind Runner case. It is a more concrete result than a glamorous trailer—but it is also much narrower.

That distinction matters because the July 25 “Claude of Duty” demo promoted as one-shotted was not one uninterrupted model pass. Its published harness used six agents per round, three rounds of parallel fan-out, blind visual critics, and then a sequential owner pass. “One-shot,” in that case, meant one human seed prompt.

Disclosure: I work with OrcaRouter and used it to run this evaluation. One OpenAI-compatible key gave me access to every model in this test, without changing how any model answered.

AI-generated illustration featuring official model logos; logos and model names are used descriptively and remain the property of their respective owners.

Want to try Claude Opus 5 yourself? Explore it on OrcaRouter.

What this one-shot actually tested

Our Opus 5 exercise was deliberately small: one model, no sub-agents, one complete brief supplied at once through Anthropic’s native messages endpoint, and a time-boxed feedback loop. The output was a Rewind Runner game, tested with a headless browser that sent fixed keys, fingerprinted changing pixels, and checked whether matching frames moved backward after a rewind. The rewind check passed 3/3 replays.

That is evidence of a specific mechanic, not a certificate of game-making greatness. The test verifies that the R-key rewind sequence responds and retraces recorded history; it does not measure whether the game is fun, visually polished, balanced, or well designed.

The viral comparison is useful mostly as a warning against compressed language. The public Claude of Duty project is a Three.js/WebGL2 FPS of roughly 55,000 lines across 11 subsystems. Its own README says it does not match modern Call of Duty, and blind reviewers selected the real game every time. That does not make the effort uninteresting. It shows why a big output, a polished video, and the word “one-shot” should not be treated as interchangeable claims.

Observation, not a leaderboard

This is n=1 per run or attempt: a case study, not a benchmark. A single question-model call is an observation, not proof of stable behavior. The human review was also limited to one named, non-blinded player for a few minutes, with no rubric. It is a play report, not evidence that a level is good—or that any challenge is impossible.

The broader controlled exercise covered GPT-5.6 Terra and Kimi K3 only: one run per model per round, four ordered prompts, vendor-default parameters, and no human quality rubric. Claude Opus 5 was tested separately, with a different setup, and must not be ranked against Terra or Kimi. Vendor effort labels should not be read as equal amounts of compute, either.

So the reader-facing takeaway is simple: ask what “one-shot” contains. Was it one prompt? One pass? A fleet of agents? Critics? Iteration? And what was actually checked afterward? A working rewind mechanic is a real, reproducible observation. It is not proof that a model can independently ship a great game.

First-party Rewind Runner test record; it documents this case study’s observed behaviour, not a universal model ranking.

Limitations

This report covers one small game, one Opus 5 case, and a behavioural mechanic check rather than a quality benchmark. It does not assess consumer-product behavior, compare API and consumer experiences, or establish stable performance across prompts, games, or runs.

Sources

First-party Rewind Runner Opus 5 case record and behavioural rewind-check record.

First-party Claude of Duty harness and project records.

This evaluation was run through OrcaRouter. The author works with OrcaRouter; model access does not imply affiliation with, endorsement by, or sponsorship from model providers. Model names and logos are used descriptively. All trademarks belong to their respective owners.