Claude Haiku 4.5 review: Sonnet 4 level smarts at Haiku speed
Claude Haiku 4.5 (anthropic-claude-haiku-4-5) is a small model that does real work. Anthropic says it matches the previous Sonnet generation (Sonnet 4) on intelligence while running 4 to 5 times faster at a much lower cost. On the hardest reasoning, and where precision matters most, the current larger models still win. The working rule: start with Haiku and move up when it falls short.
Anthropic has since released Claude Haiku 5.5 as its successor, and FastMetal serves it (model page). Haiku 4.5 is still served.
What it is
Haiku is the speed and cost tier of the Claude family. Haiku 4.5 is the first Haiku with extended thinking, computer use and context awareness.
- Context window: 200K tokens
- Maximum output: 64K tokens
That is a lot of room for a small model.
Strengths
- Coding: Anthropic reports 73.3% on SWE-bench Verified, level with the Sonnet 4 generation and very high for its size.
- Speed: about 4 to 5 times faster than the older Sonnet models, with reports of roughly 80 to 150 tokens per second and responses under 200 ms on short prompts. You notice it in live chat and pair programming.
- Sub-agents: a larger model plans and several Haiku 4.5 instances implement in parallel. Developers rate this orchestrator and worker setup as a cost-efficient design.
Weak spots
- The hardest reasoning, long careful creative writing and work that must be exactly right go better on the current Sonnet or Opus. Reviews show the bigger models clearly ahead on many benchmarks.
- Extended thinking tokens are billed as output. The point of Haiku is speed and low cost, so set how much it thinks per use case.
How it compares
| Haiku 4.5 | Bigger models (Sonnet and up) | Haiku 3.5 | |
|---|---|---|---|
| Intelligence | Reported at Sonnet 4 level | Ahead on the hardest tasks | Well behind |
| Speed | Very fast | Medium to fast | Fast |
| Extended thinking | Yes | Yes | No |
Against Haiku 3.5 it is a large step up in both intelligence and features. For its closest rival in the fast tier, see Gemini 3.5 Flash vs Claude Haiku 4.5.
Cost
Cost is the main reason to pick Haiku. It is billed per million input and output tokens at a lower rate than the Sonnet models, and at high volume that gap is your running cost. Compare the Haiku 4.5 and Sonnet 4.6 pages. FastMetal is prepaid, so you are only charged up to the balance you loaded, with no surprise bill.
FastMetal serves Haiku 4.5 in Japan, if data location matters to you.
Outside English
On routine Japanese tasks (summarizing, classifying, extraction, short replies) it gives practical quality for a small model. Nuance and natural style don't show up in benchmark numbers, though. For anything important, compare its output with a bigger model on your real prompts.
Where it fits
Good fit: summarizing, classification, tagging, first-line chat replies, real-time conversation, batch jobs, and parallel sub-agents.
Poor fit: the hardest reasoning, long careful creative writing, and work where an error is unacceptable (use Sonnet or Opus).
FAQ
Can Haiku 4.5 replace the bigger models?
No. Reviews show the bigger models ahead on many benchmarks. Use it for high-volume, low-cost work.
Is it worth moving from an older Haiku?
Usually, yes. It reaches the older Sonnet level and adds extended thinking, so it is a strong option wherever you use a small model today. Confirm on your real prompts before you switch.
How good is its Japanese?
Practical on routine tasks. Nuance is hard to measure, so test with your own prompts for anything that matters.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic-claude-haiku-4-5",
"messages": [{"role": "user", "content": "Summarize this in one sentence: ..."}]
}'
Start light tasks on Haiku 4.5 and switch model to a bigger one when the quality isn't there. Rates are on the pricing page, and every model is in the catalog.