Claude Haiku 5.5 cuts the price of a small model by up to 90 percent
Anthropic's new small model costs $0.10 per million input tokens and adds an effort dial. Independent tests put it top of its class, at a price in tokens.
- Model
- Claude Haiku 5.5
- Maker
- Anthropic
- Context
- 1M
- License
- Proprietary
- Weights
- Closed
Anthropic released Claude Haiku 5.5 on October 7. The model is better than the one it replaces, but the number that will change what people build is the price: $0.10 per million input tokens and $0.50 per million output tokens, down from $1.00 and $5.00 for Haiku 4.5.
Small models are what agents use for the boring jobs: summarizing a long log, compacting context, classifying a ticket, running as one of twenty sub-agents. A tenfold price cut changes how freely you can spend them.
How the pricing works
There is a catch in the fine print, and it is a reasonable one. The low rate applies to prompts of up to 100,000 tokens. Above that, everything costs five times as much.
| Per 1M tokens | Up to 100K prompt | Over 100K prompt | Haiku 4.5 |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache reads | $0.01 | $0.05 | $0.10 |
| Cache writes | $0.125 | $0.625 | $1.25 |
Anthropic says about 90 percent of Haiku 4.5 requests fell under the 100K line. It also says Haiku 5.5 uses a new tokenizer that produces slightly more tokens for the same text, and that after accounting for both effects the model costs around 75 percent less to run on average.
The other new control is effort. Haiku 5.5 is the first Haiku with an adjustable effort setting, with five levels from low to max, so you can decide per request how much thinking to pay for. The context window grows from 200,000 tokens to 1M.
The benchmarks
These are Anthropic’s numbers from the announcement.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| OSWorld 2.1, offline subset | 72.4% | 15.7% | 48.9% | 83.9% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 Main | 46.4% | n/a | 42.4% | 52.1% (xhigh) |
| Humanity’s Last Exam, no tools | 45.9% | 10.2% | n/a | 56.9% |
| Chartography, no tools | 46.4% | 6.4% | 29.1% | 61.6% |
| GDPval-AA v2.1 (Elo) | 1620 | 735 | 1437 | 1840 |
The jump over Haiku 4.5 is large everywhere, most of all on operating a desktop. The gap to Sonnet 5.5 is also large, and Anthropic says so: for complex agentic coding of the Terminal-Bench kind, it recommends the bigger models.
The independent view
The Decoder reports Artificial Analysis results that complicate the picture in a useful way.
- Haiku 5.5 scores 43 on the Intelligence Index, the highest of the small models. GLM-5.3 Flash is at 42, Gemini 3.8 Flash at 41, GPT-6 Luna at 38.
- That 43 is at maximum effort, where it writes about 162,000 output tokens per task, roughly three times what Luna uses. At high effort, Haiku scores 38 on about 55,000 tokens, level with Luna.
- By that analysis Luna is the more cost-efficient of the two. The last step from xhigh to max effort buys two index points for 1.8 times the tokens.
- Haiku hallucinated in 40 percent of cases on Artificial Analysis’s test, against 77 percent for Luna, though Luna answered more questions correctly overall.
In short: a higher ceiling than its main rival, and more tokens spent to reach it. The effort dial is where you choose.
Try it yourself
The API model ID is claude-haiku-5-5, and it is available on Anthropic’s platform and the major clouds. Alongside the launch, Anthropic halved cache-read pricing on Sonnet 5.5 to $0.10 per million tokens. We have not benchmarked Haiku 5.5 ourselves: this is based on Anthropic’s announcement and The Decoder’s reporting.
Disclosure: Claude, an Anthropic model, helps write this site.