A clean sweep makes a good headline. Before using it to choose an AI model, look at the finish line: this contest stops after each side has played just 25 moves and awards the result by remaining material.
A result with a specific job description
In an original experiment, JC BuenaVentura pits GPT-6 Astra against Gemini 4 Argon across four chess games. The complete creator captions supplied for this article report four Astra wins. We have not rerun the games or independently verified their scores. The experiment deserves its own attribution: https://www.youtube.com/watch?v=Wq5SmovcgR0
| Game | Astra’s colour | Astra material | Argon material |
|---|---|---|---|
| 1 | White | 16 | 9 |
| 2 | Black | 14 | 13 |
| 3 | White | 20 | 18 |
| 4 | Black | 19 | 13 |
| Total | — | 69 | 53 |
These are the creator’s reported material scores. They are not Elo ratings, and four games do not establish either a dependable chess rating or an all-purpose model ranking.
The contest values queens at nine points, rooks at five, bishops and knights at three, and pawns at one. Checkmate scores 39–0. Otherwise, play ends at 50 plies—a ply is one side’s move—and the remaining material decides the winner. Equal material counts as a loss for both models.
That cutoff is different from FIDE’s fifty-move rule. Under Article 9.3, a valid draw claim concerns fifty moves by each player without a pawn move or capture. It is not a deadline twenty-five moves into a game. Source: https://handbook.fide.com/chapter/e012023
The deadline becomes part of the strategy
An even-ply cutoff lets Black act last. A capture on that final turn cannot receive an immediate reply. This follows from the format; it does not prove that the last-move opportunity caused any particular result. Colours alternate in the experiment, and Astra reportedly wins with both.
The interesting question is therefore narrower than “Which AI is smarter?” Can a model track the board, protect material and choose an action consistent with the stated objective? A move that helps win this short material contest may differ from the best route to winning a full game.
The creator also describes cases where an explanation and the submitted move diverge. Those are useful leads for an audit, not something this article independently reproduces. A complete game record and raw responses would let a reader compare the stated plan with the action. The supplied captions do not include those machine-readable records.
Matching labels is not matching computation
The model names are documented: OpenAI lists gpt-6-astra and Google lists gemini-4-argon. Both document a medium reasoning setting. A shared setting name does not establish an identical amount of computation. OpenAI: https://developers.openai.com/api/docs/models/gpt-6-astra ; Google: https://ai.google.dev/gemini-api/docs/latest-model?hl=en
The creator says both ran at the same temperature and with a 32,000-token ceiling per move. Tokens are the small pieces of text processed by a model; a ceiling is not actual consumption. There is also a concrete qualification: Google’s current Argon guide says temperature, top_p and top_k are ignored if supplied. “Same temperature” therefore cannot, by itself, establish matched sampling behaviour. The guide does not reveal which requests or provider route this experiment used; raw logs are needed for that. Source: the Google guide’s migration section, linked above.
The bill needs its own evidence
The creator reports spending $6.35 on Astra and $0.98 on Argon, with roughly twelve versus six seconds per move. Calculated from those reported amounts, the spending ratio is about 6.48 and the difference is $5.37. These are observations claimed for one run, not general speed or cost guarantees.
The official standard API text rates checked on 3 November are $10 input/$50 output per million tokens for Astra. Argon lists introductory rates of $2/$10 through 30 November 2026, then $4/$20 from 1 December. These are uncached standard rates; Astra has additional long-input and service-tier conditions. Pricing sources: https://developers.openai.com/api/docs/pricing and https://aistudio.google.com/docs/pricing#gemini-4-argon . Neither rate card verifies the video’s totals without actual usage, caching, routing and the prices in force when the run occurred.
Another claim deserves the same care: the creator reports zero reasoning tokens on 61 of Argon’s 100 moves and nine of Astra’s. OpenRouter documents differences in reasoning visibility across models and providers. We have not audited these usage records; a reported zero alone does not establish that no internal computation occurred. Source: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens
Milo’s assessment: Keep the action in view
For model selection, the most transferable lesson is the gap between a plausible explanation and a useful action. A system can sound organised while choosing a move that fails its own plan. In your workflow, judge the action separately from the prose around it.
A stronger chess comparison would replay shared positions with reversed colours and repeated trials. Report this short material format separately from standard games. Keep exact model and provider IDs, accepted settings, raw outputs, invalid moves, retries, usage charges and elapsed time. FEN records board positions; PGN records games. Publishing those records would make the claims easier to check. This is a proposed evaluation plan, not a test we have performed.
The reported sweep answers who won these four contests. Choosing a model for your work requires a different answer: which system reaches the accuracy you need at an acceptable total cost and delay?
