All articles

SKILL.state: Input, Cache, and Output

This note follows the first article on the SKILL.state experiments, which describes the protocols, run conditions, and criteria for full success.

The data in this note come from the Codex series run on September 8–9: two models, four modes, five projects, and five repetitions of each project. Each row covers 25 sessions; the totals include timeouts and failed runs.

Input counts tokens in the main agent loop, including cached tokens. “Cached share” is the proportion of these tokens that the provider classified as cached input. Output counts tokens generated by the model in the main agent loop.

Model Mode Input, M tokens Cached share Uncached input, M tokens Output, K tokens Full success
Sol Native 6.367 89.4% 0.673 270.8 24/25
Sol Paper 9.215 78.9% 1.946 402.7 19/25
Sol V2 4.077 52.3% 1.943 221.4 23/25
Sol V3 4.180 47.2% 2.207 251.9 25/25
Astra Native 2.734 85.3% 0.402 150.0 25/25
Astra Paper 15.318 75.0% 3.833 324.5 16/25
Astra V2 2.654 63.7% 0.964 133.6 25/25
Astra V3 2.340 56.5% 1.018 130.5 25/25

Modes: Native — the standard agent loop; Paper — the first adaptation of SKILL.state with a single text observation; V2 — structured state and a three-observation window; V3 — V2 with batches of actions executed sequentially.

What the numbers show

The cached share describes the composition of input tokens, not the proportion of requests that hit the cache. It does not equal financial savings: calculating cost requires separate rates for uncached input, cached input, and output. Output also does not measure the quality or amount of the final code.

Source