Claude Haiku 5.5
Claude Haiku 5.5 is Anthropic's October 7, 2026 model for high-volume tasks. Use claude-haiku-5-5: 1M context, 128K maximum output, text/image input and text output. API rates rise when prompt length exceeds 100,000 tokens.
A low-cost model is useful when many narrow decisions dominate a workflow. Compare accuracy, correction work and total cost at the relevant effort level; a larger context does not imply a flat rate.
Official launch and current provider documentation support identity, access and specification claims. Publisher performance claims are attributed; no independent model call or task-quality test was run.
- Classification and routing
- Document extraction and short summaries
- Bounded subagent work
- Latency-sensitive assistance
For prompts up to 100,000 tokens, Claude API rates per million are $0.10 input and $0.50 output; above 100K they become $0.50/$2.50. Cache reads are $0.01/$0.05, five-minute writes $0.125/$0.625 and one-hour writes $0.20/$1 at the same threshold. Count the new tokenizer's actual tokens before budgeting.
Adaptive thinking defaults to medium effort. Responses can begin with thinking blocks, so select answer blocks by type. Ordinary maximum output is 128K; the separate Message Batches beta can reach 300K with its required header. Thinking counts toward max_tokens, so a very small cap can exhaust the budget before answer text.
Remove manual budget_tokens and sampling parameters, avoid assistant prefill and check the current computer/browser tool versions. Bedrock structured-output support differs from Claude API. Earlier thinking blocks remain in context and are bound to the producing or linked account. Anthropic documents these breaking and behavioral changes; replay representative conversations before migration.
Summaries, routing, extraction and short subagent jobs are useful trials. The launch discussion asks whether Haiku can be a main coding agent; customer quotes and publisher benchmarks do not settle that for a repository. Compare fixed tasks, effort and tools with Sonnet/Opus, counting refusals, rework and accepted results.
Claude Haiku 5.5 FAQ
Common questions about Claude Haiku 5.5.
Does Haiku 5.5 always cost $0.10/$0.50 per million tokens?+
No. Those rates apply to prompts up to 100K tokens. Above 100K, input/output rates are $0.50/$2.50, with cache rates also increasing.
Can Haiku 5.5 generate images?+
No. It accepts text and images and returns text. Browser/computer tools can act in an external environment; that does not create native image output.
Is switching from Haiku 4.5 just a model-name change?+
No. Adaptive thinking, tokenization, sampling, prefill, computer tools and platform-specific structured outputs need migration checks.