Efficiency
Tokens and calls as reported by each agent per step (the gym records what the client sends). Figures are per agent-episode, averaged over the ranked scenarios. Input tokens include prompt-cache reads and writes; cost uses the prices given to the build (none were given, so cost is not shown), with cache reads at 0.1× and writes at 1.25× the input price unless stated.
Usage and cost per episode
| Entry | Index | Calls | Input tokens | Cached shareof input | Output tokens | Tokens / sim day | Cost | PricesUSD per M in / out |
|---|---|---|---|---|---|---|---|---|
| builtin:rules baseline | 82.4 | — | — | — | — | — | — | — |
| builtin:idle baseline | 49.1 | — | — | — | — | — | — | — |
| builtin:random baseline | 24.1 | — | — | — | — | — | — | — |
Baselines and human players make no model calls. Wall-clock time is not compared: it depends on the machine and the provider.