Claude Haiku 4.5: what it costs to run
The price Anthropic lists for Claude Haiku 4.5, what a support reply, a document summary and an agent task cost at it, and where that puts it among 37 models.
US dollars per million tokens, standard rate, shortest context band.
- 1,000 support replies
- $5.00
- Cached input
- $0.10
- Context window
- 200,000 tokens
- Released
- 15 Oct 2025
- Weights
- Closed
From Anthropic’s own pages, 22 September 2026.
Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens on Anthropic’s own API, and $0.10 for cached input. At those prices 1,000 support replies cost $5.00, the 17th most expensive of 33 priced models on our list.
What a job costs
Three jobs, 1,000 runs of each, at the list price above, and where each puts Claude Haiku 4.5 among the 33 models with a list price.
| Job | 1,000 runs | Among 33 |
|---|---|---|
| Support reply3,000 in, 400 out. | $5.00 | 17th most expensive |
| Document summary25,000 in, 1,000 out. | $30.00 | 16th cheapest |
| Agent task80,000 in (70,000 cached), 3,000 out. | $32.00 | 14th cheapest |
The token counts are ours, chosen to look like real work; the prices are Anthropic’s. All three jobs stay under 200,000 tokens a request, below every long-context surcharge on the list.
Claude Haiku 4.5 on real small-business tasks
What the AIGROW benchmark measured when Claude Haiku 4.5 was given 20 office tasks, each marked pass or fail by fixed checks.
- Success rate
- 35% of 20 runs passed (95% interval 18 to 57%), 32nd of 36 models.
- Cost per 1,000 passes
- $6.62, failed runs included.
- Answer time
- 2.5 s median, 3.4 s at the 90th percentile.
- Strongest and weakest
- Email triage (100% passed) and review replies (0%).
The small print
What the base rate above leaves out, copied from the maker’s pages: surcharges, discounts, promotions and cache pricing.
Cache write $1.25 (5 min) or $2 (1 h). Batch API 50% off.
Claude Haiku 4.5 against the models it is weighed with
Each comparison puts both list prices side by side and costs the same three jobs on each.
- vs GPT-5.6 Luna
- Anthropic's small model costs five times Luna on input, so it has to earn the difference.
- vs Gemini 3.8 Flash
- Fast models from Anthropic and Google weighed for customer-facing replies, with Flash cheaper on both input and output.
- vs GPT-5.4 mini
- Anthropic's current small model against OpenAI's previous mini, at similar list prices.
Other Anthropic models
| Model | Input | Output | Context | 1,000 replies |
|---|---|---|---|---|
| Claude Sonnet 5Anthropic | $2.00 | $10.00 | 1M | $10.00 |
| Claude Sonnet 4.6Anthropic | $3.00 | $15.00 | 1M | $15.00 |
| Claude Opus 5Anthropic | $5.00 | $25.00 | 1M | $25.00 |
| Claude Fable 5.1Anthropic | $10.00 | $50.00 | 1M | $50.00 |
Where the price comes from
Read on 22 September 2026. Each price is the standard rate for the shortest context band, with no batch or priority discount; long-context surcharges, batch rates and promotions are in each model’s small print. No price was taken from a reseller.
Pay a quarter of list price
AIGROW API credit is metered at the makers’ list prices, and $25 buys $100 of usage. The credit page lists the models it covers; ask about any other before buying.
Prices checked 22 September 2026