TL;DR — Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens — Anthropic's budget tier, roughly 5× cheaper than Opus 5 and the cheapest way to get Claude-family quality. Cache hits at ~10% of input and the 50% Batch API discount both apply. It is priced above the open-weight bulk tier (DeepSeek Flash, Gemini Flash) but delivers noticeably stronger instruction-following, which is exactly the trade a router should be making case by case.
Official pricing (as of August 2026)
| Input /1M | Output /1M | |
|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 |
| Cache hits | ~10% of input rate | — |
| Batch API | $0.50 (-50%) | $2.50 (-50%) |
As across the Anthropic lineup, batch and cache discounts stack — high-volume Haiku pipelines with stable prompts can run at a fraction of headline rates.
What this costs in practice
- Chat product (10K requests/day, ~800 in / 400 out): ~$84/month
- Summarization pipeline (50M in / 10M out monthly): $100/month, or ~$50 batched
- Agent background tasks: usually better routed to Flash-class models unless Claude-specific behavior matters
How it compares
| Model | Input /1M | Output /1M | Note |
|---|---|---|---|
| Claude Sonnet 5 | $2.00 | $10.00 | 2× Haiku, and now a permanent rate |
| Claude Haiku 4.5 | $1.00 | $5.00 | This page |
| GPT-4o-mini | $0.15 | $0.60 | Details |
| DeepSeek V4 Flash | $0.14 | $0.28 | Open-weight bulk tier |
Where it fits in a routing strategy
Haiku 4.5 is the classic 'middle seat' — too expensive for pure bulk work, strong enough to rescue tasks Flash-class models fumble. Routers earn their keep here: ClawRouters escalates to Haiku when a cheap model's output fails validation, keeping the average cost near bulk-tier while capping the quality floor. See which providers let you mix tiers.
Routing decision card
The exclusive layer: the "middle seat" economics no rate card explains.
- Route here when: a bulk-tier model's output failed validation and the task needs Claude-family instruction-following without mid-tier prices — Haiku is the rescue tier, not the default tier
- Do not route here when: the task is pure bulk (Flash-class wins on price 7-15×) or genuinely hard (Haiku retries on hard reasoning waste more than they save)
- Head-to-head at this price point: nothing sits exactly at $1/$5 — that isolation is Haiku's niche: the cheapest "definitely competent" option
- Typical pairing: Flash-first with Haiku fallback keeps average cost near bulk while capping the quality floor
Prices from the provider's official rate card as of August 2026 — verify at the source before committing to volume, as rates change with model launches. Full market context: 2026 AI pricing guide · routing basics.