Thinking level, chess results, and cost⁠

More thinking does not guarantee more wins

A higher reasoning setting can increase request time and output tokens. It does not guarantee a better move or a higher benchmark rating. Compare the measured results for each saved setup instead of assuming that the highest level is strongest.

Thinking labels are provider settings, not a shared unit of computation. Medium on one model need not use the same work as Medium on another.

Read four measures together

Use the rating and win, draw, and loss record to understand observed results. Use time per turn to understand the model’s request time. Use cost per game and output tokens per game to understand resource use.

Time per turn includes model request time per accepted move. Engine analysis and video rendering are separate. Output totals include reasoning when the provider reports it. Average game cost uses completed benchmark games with complete pricing; N/A means a complete average is unavailable.

A current example: Gemini 4 Argon

The table below reads the current saved Argon Medium and High results. It includes the sample count and update time so the comparison can be checked. Open the model page for the visible game records.

Both setups played the placement pool with both colours. This is an observed benchmark comparison, not a controlled claim about all reasoning tasks. Changes in the opponent pool and later results can change both ratings.

Live results updated 2026-10-26 UTC. Gemini 4 Argon match evidence.

Choose a human opponent from the current offer

Benchmark cost is not a quote for a player’s OpenRouter bill. Models marked BYOK use the signed-in player’s key and actual provider charges. A temporary FREE offer changes human-game billing, not the benchmark settings or ratings.

Check the live challenge panel for current availability and any offer deadline. Choose a level on measured strength, response time, and the billing mode that is shown before starting a game.