Claude Haiku 5.5: The Small Model Built for Work at Scale
Anthropic released Claude Haiku 5.5 on October 7, completing the 5.5 lineup after Opus 5.5 and Sonnet 5.5. Haiku is the model to look at when a system needs to make many small decisions quickly: classify a request, summarize a document, look up a fact in a database, or send a narrowly scoped subtask to an agent.
The launch is about cost per useful result. Anthropic says Haiku 5.5 is its fastest model at standard speed and costs about 75% less per task on average than Haiku 4.5. That figure combines a sharp price cut with changes in how many tokens the model uses. It is a vendor estimate, so the real saving depends on your prompts, output length, and error rate.
What changed
- Much lower prices for short prompts. For prompts up to 100,000 tokens, input costs $0.10 and output costs $0.50 per million tokens. That is one tenth of Haiku 4.5’s input and output prices.
- A higher price above 100,000 tokens. Input rises to $0.50 and output to $2.50 per million tokens. Both are still half the corresponding Haiku 4.5 prices.
- Adjustable effort. This is the first Haiku model with an effort setting. You can spend less on simple requests and allow more thinking for harder ones.
- Stronger computer use and reasoning. Anthropic reports large gains over Haiku 4.5, while Sonnet 5.5 remains stronger on complex, open-ended tasks.
Anthropic says roughly 90% of requests to Haiku 4.5 had prompts below the 100,000-token threshold. That explains the aggressive short-prompt pricing. It also means a long-context workflow needs its own cost calculation; the headline rate will not describe every call.
The benchmark gains are large, with a clear ceiling
On Anthropic’s published evaluations, Haiku 5.5 moves well beyond Haiku 4.5:
| Evaluation | Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| OSWorld 2.1, offline subset | 72.4% | 15.7% | 83.9% |
| Humanity’s Last Exam, with tools | 57.4% | 18.7% | 64.5% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 70.6% |
| FrontierCode 1.1, main | 46.4% | — | 52.1% |
These are Anthropic’s launch results, not a promise that every application will see the same jump. The OSWorld result is particularly interesting for browser and computer agents: Haiku 5.5 reaches 72.4% on the offline subset, versus 15.7% for Haiku 4.5. Yet the Terminal-Bench gap to Sonnet remains wide. A small model can handle more tool-driven work than before, but long, messy coding jobs still favor a larger one.
That suggests a practical division of labor. Let Haiku summarize logs, extract facts, triage issues, and perform bounded browser steps. Use Sonnet or Opus to decide what to build, resolve ambiguous failures, and review work that crosses many files or systems. Anthropic specifically describes Haiku as a subagent paired with the larger models.
Token prices are only part of the bill
The published prices per million tokens are:
| Token type | Haiku 5.5, prompt up to 100k | Haiku 5.5, prompt over 100k | Haiku 4.5 |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache read | $0.01 | $0.05 | $0.10 |
| Cache write | $0.125 | $0.625 | $1.25 |
The 90% per-token reduction on shorter prompts does not mean every workload becomes 90% cheaper. Haiku 5.5 uses a newer tokenizer. Anthropic’s migration guide says the same input text produces approximately 30% more tokens than on Haiku 4.5, although the exact change varies by content. Some tasks may also use more thinking or require different output lengths. Measure the full task: total tokens, elapsed time, retries, and whether a person had to correct the answer.
One related announcement improves the larger-model side of that calculation. Anthropic halved Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per million tokens and estimates this makes most agentic Sonnet work about 20% cheaper. If you already route between Haiku and Sonnet, recalculate both branches rather than moving everything to Haiku by default.
What developers should check before switching
The API model ID is claude-haiku-5-5, and Anthropic says the model is available on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. The migration is more than a name change for some integrations:
- Recount tokens and revisit limits. The newer tokenizer can make a
max_tokensvalue tuned for Haiku 4.5 too small for an equivalent answer. - Replace fixed thinking budgets. Requests using
thinking: {"type": "enabled", "budget_tokens": ...}need adaptive thinking. Useoutput_config.effortto tune how much work the model does. - Check response parsing. Thinking blocks can appear before text, so read content blocks by
typeinstead of assuming the first block is the answer. - Review older request patterns. The migration guide lists changes to sampling parameters, assistant prefills, and computer-use tools that can otherwise cause API errors.
For a production trial, start with a set of real high-volume tasks and compare Haiku 4.5, Haiku 5.5 at several effort settings, and Sonnet 5.5. Track accuracy, false positives, p95 latency, total cost per completed task, and the rate at which work must be escalated to a larger model. A low token price matters only if the work still gets done correctly.
The practical takeaway
Haiku 5.5 makes the small-model tier more useful for agents, not just cheaper for chat. Its strongest case is frequent, bounded work where speed and price compound across thousands or millions of calls. The launch benchmarks support trying it for extraction, triage, summarization, and computer-use subtasks. They also show why Sonnet and Opus still belong in the route for difficult coding and decisions that need sustained judgment.
The simplest next step is to replay your own workload, especially prompts near the 100,000-token boundary. Tune effort and compare cost per correct outcome. That will tell you where Haiku 5.5 earns its place.