Claude Haiku vs Claude Sonnet
On a typical agent workload, Claude Haiku runs about 73% cheaper than Claude Sonnet at list price ($1,567 against $5,870 a month). Here is the full cost breakdown.
| metric | Claude HaikuAnthropic · Cheap tier | Claude SonnetAnthropic · Mid tier |
|---|---|---|
| Input priceper 1M tokens | $0.80 | $3.00 |
| Output priceper 1M tokens | $4.00 | $15.00 |
| Blended price3:1 input:output mix, per 1M | $1.60 | $6.00 |
| One chat request1.5k in / 600 out, no tools | $0.0036 | $0.01 |
| Agent workload / month2,000 req/day, 3 tool calls, RAG on | $1,567 | $5,870 |
cheaper · public list prices as of 2026-06 · estimates, not quotes
which_to_choose
Which one should you pick?
On price alone, Claude Haiku wins. It comes in around 73% cheaper than Claude Sonnet on the same agent workload ($1,567 against $5,870 a month at 2,000 requests a day), and the gap widens as volume and tool calls grow, because every tool call re-sends the context and you pay for it at each model's rate.
The case for Claude Sonnet comes down to fit. If it resolves a task in fewer attempts or shorter prompts on your workload, the higher per-token rate can still come out ahead of a cheaper model that needs retries. Price the two on your own evaluation set and your actual token mix before you commit, because the list price rarely decides it alone.
Claude Haiku: Anthropic's fastest, lowest-cost tier, aimed at high-volume and simpler tasks. Claude Sonnet: Anthropic's balanced model and a common default for production agents.
faq
Questions & answers
- Is Claude Haiku or Claude Sonnet cheaper?
- Claude Haiku is cheaper at list price. It runs $0.80 per million input tokens and $4.00 per million output tokens, against $3.00 and $15.00 for Claude Sonnet. On a typical agent workload that works out to about 73% less per month.
- What is the price difference between Claude Haiku and Claude Sonnet?
- Claude Haiku is $0.80 in and $4.00 out per million tokens; Claude Sonnet is $3.00 in and $15.00 out. Output tokens cost several times more than input on both, so the gap that matters most depends on how much your workload generates versus reads.
- Should I switch from Claude Sonnet to Claude Haiku to cut cost?
- Possibly. Claude Haiku is about 73% cheaper on the same workload, and the saving grows with volume and tool calls because each tool call re-sends the context. But a cheaper model that needs retries or longer prompts can cost more in practice, so price both on your own evaluation set and your actual token mix before you switch.
Picking a model is the easy part. Making it cheap in production is the work.
Prompt caching, context trimming, and the right tier per task usually cut an LLM bill by more than switching models. Book a call, or leave your email and I'll reach out.
Prefer proof first? See how this plays out in real case studies →