Claude Haiku 4.5 in Claude Code: tokens per question
Haiku is the lightest Claude tier in this comparison, suited to focused coding questions where a short answer is enough.
The figures
- Bare Claude Code
- 55,500
- With capsul
- 19,300
- less input sent
- x2.88
- answer checks passed out of five, bare then with capsul
- 2 / 4
- tool steps per question, bare then with capsul
- 2.4 / 4.4
- output tokens per question, bare then with capsul
- 2,710 / 444
- API-equivalent cost per question, bare then with capsul
- $0.0812 / $0.0265
cache-weighted input tokens per question
cache-weighted input tokens per question
90% interval: x1.62 to x4.49
Its API list price is the lowest in the Claude group. That makes it a natural candidate for locating a function, explaining a test failure or making a small, well-scoped edit. A cheap model can still spend heavily if the agent reads many files before answering, so compare the input column rather than the prompt alone.
The measured interval matters particularly here. This row should not be read as proof that capsul saves input on every Haiku question. Look at the question-level results and answer checks before choosing it for a workflow that repeats often.
Question by question
Five one-line questions about a real TypeScript codebase, each asked 3 times per arm. Weighted input per question, mean of the repetitions.
| Question | Bare | capsul | Less input | Correct runs, bare / capsul |
|---|---|---|---|---|
| Q1 | 57,700 | 18,000 | x3.20 | 1 / 1 |
| Q2 | 53,600 | 24,800 | x2.16 | 1 / 3 |
| Q3 | 18,200 | 10,700 | x1.70 | 0 / 3 |
| Q4 | 118,000 | 30,600 | x3.87 | 3 / 3 |
| Q5 | 29,700 | 12,200 | x2.43 | 3 / 3 |
List prices
| Model | Input | Output | Cache read | Cache write 5 min | Cache write 1 h | Context |
|---|---|---|---|---|---|---|
| Claude Haiku 4.5 | 1.00 | 5.00 | 0.10 | 1.25 | 2.00 | 200,000 |
| Claude Sonnet 5 | 2.00 | 10.00 | 0.20 | 2.50 | 4.00 | 1,000,000 |
| Claude Sonnet 4.6 | 3.00 | 15.00 | 0.30 | 3.75 | 6.00 | 1,000,000 |
| Claude Opus 4.6 | 5.00 | 25.00 | 0.50 | 6.25 | 10.00 | 1,000,000 |
| Claude Opus 4.7 | 5.00 | 25.00 | 0.50 | 6.25 | 10.00 | 1,000,000 |
| Claude Opus 4.8 | 5.00 | 25.00 | 0.50 | 6.25 | 10.00 | 1,000,000 |
| Claude Opus 5 | 5.00 | 25.00 | 0.50 | 6.25 | 10.00 | 1,000,000 |
| Claude Opus 5.5 | 4.00 | 20.00 | 0.20 | 5.00 | 8.00 | 1,000,000 |
| Claude Fable 5 | 10.00 | 50.00 | 1.00 | 12.50 | 20.00 | 1,000,000 |
| Claude Fable 5.1 | 10.00 | 50.00 | 0.25 | 12.50 | 20.00 | 1,000,000 |
Anthropic list prices, checked September 23, 2026.
How this was measured
Two arms on the same model: Claude Code as anyone runs it, and the same question through capsul. The unit is cache-weighted input: fresh tokens at full weight, cache reads at a tenth. Campaign of September 22, 2026, 3 repetitions per question and arm.
30 / 30 cells · /benchmarks/2026-09-22-claude.json
Questions
How many tokens does Claude Haiku 4.5 use per question in Claude Code?
In the September 22, 2026 benchmark, a bare Claude Code session on Claude Haiku 4.5 sent 55,500 cache-weighted input tokens per one-line coding question on average. With capsul on the same model, 19,300.
How much does a coding question cost on Claude Haiku 4.5?
At Anthropic list prices, the bare agent's traffic came to $0.0812 per question and capsul's to $0.0265. A Claude subscription does not bill per token: read it as how much of a plan each question uses, compared with other models.
Is the saving on Claude Haiku 4.5 established?
Yes. The 90% interval runs from x1.62 to x4.49, entirely above one, on 30 measured cells.
Does capsul change Claude Haiku 4.5's answers?
It passed more of them: 2 of five answer checks bare, 4 of five with capsul. Five checks is a small sample: read it as no loss rather than a gain.