/ capsul models
Every model, measured
What each model sends per question in Claude Code and Codex CLI, with and without capsul. Every figure comes from a dataset you can download.
Cache-weighted input tokens per one-line coding question, mean of the repetitions. Computed from the published datasets: Claude campaign of September 22, 2026, Codex campaign of September 4, 2026.
Claude models, in Claude Code
| Model | Bare agent | With capsul | Less input | Answer checks |
|---|---|---|---|---|
| Claude Haiku 4.5 | 55,500 | 19,300 | x2.88 | 2 / 4 |
| Claude Sonnet 5 | 54,200 | 18,900 | x2.86 | 4 / 4 |
| Claude Sonnet 4.6 | 31,400 | 16,600 | x1.89 | 4 / 4 |
| Claude Opus 4.6 | 45,700 | 16,400 | x2.78 | 4 / 4 |
| Claude Opus 4.7 | 38,700 | 20,100 | x1.92 | 5 / 4 |
| Claude Opus 4.8 | 65,800 | 13,800 | x4.75 | 5 / 4 |
| Claude Opus 5 | 45,300 | 16,200 | x2.80 | 4 / 4 |
| Claude Opus 5.5 | 42,300 | 15,900 | x2.66 | 3 / 4 |
| Claude Fable 5 | 41,400 | 17,600 | x2.35 | 4 / 4 |
| Claude Fable 5.1 | 49,700 | 23,400 | x2.12 | 3 / 4 |
GPT models, in Codex CLI
| Model | Bare agent | With capsul | Less input | Answer checks |
|---|---|---|---|---|
| gpt-5.6-sol | 33,100 | 14,700 | x2.26 | 3 / 3 |
| gpt-5.6-terra | 24,000 | 11,100 | x2.16 | 3 / 4 |
| gpt-5.6-luna | 32,100 | 13,600 | x2.36 | 4 / 3 |
| gpt-5.5 | 47,200 | 8,130 | x5.81 | 4 / 4 |
| gpt-5.4 | 31,800 | 8,220 | x3.86 | 4 / 4 |
| gpt-5.4-mini | 42,300 | 12,100 | x3.49 | 3 / 4 |