Claude Haiku 5.5 review

Claude Haiku 5.5 is Anthropic's new high-volume model for classification, extraction, routing, summaries, compaction, browser use, and small subagent steps. It launched on October 7, 2026 with a claude-haiku-5-5 API ID, a 1M-token context window, adaptive thinking, and a lower rate card than Haiku 4.5.

The short verdict is conditional: Haiku 5.5 is a strong default for narrow, repeatable work where latency and cost matter. It is not a drop-in replacement for a larger model on open-ended architecture, difficult debugging, or long agentic coding. The useful upgrade is the combination of much lower token prices, the first adjustable effort setting in the Haiku line, and materially higher published scores than Haiku 4.5.

Claude Haiku 5.5 at a glance

Anthropic describes Haiku 5.5 as its fastest model at standard speed and its most capable small model so far. The platform documentation positions it for latency-sensitive, high-volume work. It supports the same broad text-and-image input shape as the larger Claude 5.5 models, but the model's value comes from doing a bounded job quickly rather than carrying every difficult decision.

The 1M context window is a capacity limit, not included usage. The 128K limit is the normal synchronous maximum; the 300K figure belongs to the Message Batches API beta path. Keep those contracts separate when you size a client or compare costs.

What improved over Haiku 4.5?

Anthropic's launch page reports the following results. The values are useful signals about the intended tier, not a universal ranking of every production task.

The largest practical story is not that Haiku suddenly becomes an Opus-class coding model. Anthropic's own launch material says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding. Haiku 5.5 is aimed at the smaller steps around that work: extracting a value from a document, classifying a ticket, compacting context, checking a browser page, or acting as a subagent while a larger model owns the plan.

The benchmark table also mixes different harnesses, tools, effort settings, and scoring rules. A 39.2% Terminal-Bench result does not tell you the pass rate in your repository. Treat it as a reason to test a narrow coding task, not as permission to route every coding task to the cheapest model.

Claude Haiku 5.5 pricing and task cost

The direct Claude API uses two prompt-size bands. For a prompt up to 100K tokens, the input rate is $0.10 and output is $0.50 per million tokens. Above 100K, the published input and output rates are $0.50 and $2.50. Cache and Batch pricing use the same bands.

For a small uncached request with 100K input tokens and 20K output tokens, the direct cost is 0.1 × $0.10 + 0.02 × $0.50 = $0.02. The same token shape on Haiku 4.5 is 0.1 × $1 + 0.02 × $5 = $0.20, a 90% lower rate-card result before accounting for different token counts. Anthropic says Haiku 5.5 uses its newer tokenizer, which produces approximately 30% more tokens for the same text, and reports around 75% lower average task cost rather than promising a universal 90% saving.

For a 200K-input request with 20K output, the over-100K calculation is 0.2 × $0.50 + 0.02 × $2.50 = $0.15. A 100K cache read plus 20K output is 0.1 × $0.01 + 0.02 × $0.50 = $0.011, excluding the earlier cache-write charge. These are arithmetic examples, not measured bills; thinking tokens, tools, retries, cloud-provider premiums, and data-residency multipliers can change the total.

OmniaKey has its own live model catalog and gateway quote. Check the current model catalog before assuming that a newly released Anthropic model is already callable through the gateway. The Claude versus OpenAI pricing comparison explains why direct-provider rates and gateway rates should remain separate.

Where Haiku 5.5 fits

This routing rule is a starting point, not a benchmark result. Compare models on the same prompt, tools, permissions, effort, acceptance test, and retry budget. Count the cost of failed attempts and human correction, then divide total spend by accepted work.

Migration from Haiku 4.5 can break a working client

The model ID change is easy. The request-contract changes are where a production integration can fail.

1. Use the fixed ID. Set claude-haiku-5-5. There is no date-suffixed alias in the current Claude API contract.
2. Recount tokens. The newer tokenizer produces about 30% more tokens for the same text, so revisit max_tokens, quota alerts, and cost estimates.
3. Move to adaptive thinking. thinking: {"type":"enabled","budget_tokens":N} returns a 400 error. Leave thinking unset or send {"type":"adaptive"} and use output_config.effort to control depth.
4. Parse blocks by type. Adaptive thinking can add thinking blocks before text. The default block can contain an empty thinking field and a signature; do not assume the first block is the answer.
5. Remove sampling knobs. Omit temperature, top_p, and top_k. Non-default values are rejected, and keeping old client defaults can create a 400 response.
6. Remove assistant prefills. Haiku 5.5 rejects a final assistant turn used as a prefill. End the request with a user turn and use structured outputs or tools for constrained formats.
7. Update computer use. On the Claude API and Google Cloud, replace computer_20250124 with computer_toolset_20260801 and update the tool loop.
8. Keep thinking conversations append-only. Replay thinking blocks through the account that produced them and do not edit earlier system, tool, or message content before a bound block.
9. Handle refusals explicitly. A safety refusal can return stop_reason: "refusal"; Haiku 5.5 does not provide a server-side fallback.
10. Check capacity commitments. Priority Tier commitments for Haiku 4.5 do not automatically carry over to Haiku 5.5.

The smallest adaptive request shape is:

{
  "model": "claude-haiku-5-5",
  "max_tokens": 4096,
  "thinking": {"type": "adaptive"},
  "output_config": {"effort": "medium"},
  "messages": [
    {"role": "user", "content": "Extract the three risks and return JSON."}
  ]
}

For Claude Code-specific setup, use the Claude Code API setup guide. Keep the route and model ID explicit, and check the Sonnet 5.5 review when the task needs a larger daily model.

A fair evaluation before you switch

We did not run a paid Haiku 5.5 benchmark for this review. A useful pilot can still be small and reproducible:

Run each case with a fresh conversation, fixed tools and permissions, explicit effort, and the same acceptance command. Report the sample size and failures. If you compare Haiku 5.5 with Sonnet or Opus, keep effort settings and provider routes in separate tables. The goal is accepted-task cost, not a leaderboard built from one impressive answer.

Verdict

Claude Haiku 5.5 is the most compelling small-model upgrade in this release cycle for high-volume, latency-sensitive work. Its published improvements over Haiku 4.5 are substantial, the under-100K rate card is unusually low, and adjustable effort makes the cost-quality trade-off easier to control.

Use it as the fast layer around a larger agent: classify, extract, summarize, compact, browse, and handle bounded subagent work. Keep Sonnet 5.5 or Opus 5.5 for architecture, difficult debugging, and decisions where a wrong answer costs more than a few extra tokens. Migrate deliberately, because tokenizer changes, adaptive thinking, prefills, toolsets, and block parsing can turn a model swap into a production bug.

Frequently asked questions

When was Claude Haiku 5.5 released?

Anthropic released Claude Haiku 5.5 on October 7, 2026. The current Claude API model ID is claude-haiku-5-5.

How much does Claude Haiku 5.5 cost?

For prompts up to 100K tokens, direct API input is $0.10 per million and output is $0.50 per million. Larger prompts use the published $0.50 / $2.50 rates. Cache reads are $0.01 / $0.05 by prompt tier.

Is Haiku 5.5 better than Haiku 4.5?

Anthropic's published scores are higher across the listed knowledge-work, computer-use, reasoning, coding, and visual tests. They are vendor results under specific harnesses, so validate your own acceptance tasks before changing production routing.

Is Haiku 5.5 good for coding?

It is a good fit for narrow coding subagent steps, extraction, test triage, and mechanical edits with clear checks. Anthropic still positions Sonnet 5.5 and Opus 5.5 as better choices for complex, open-ended agentic coding.

Can I reuse Haiku 4.5 thinking settings?

No. Replace manual enabled thinking with adaptive thinking and an effort setting. Also recount tokens, remove old sampling parameters, parse response blocks by type, and remove assistant prefills.

Is Claude Haiku 5.5 available on OmniaKey?

Availability can change independently of Anthropic's launch. Check the live OmniaKey model catalog and use the exact ID shown there; this review does not claim a Haiku 5.5 gateway route.

Sources checked

• Anthropic: Introducing Claude Haiku 5.5
• Claude Haiku 5.5 model overview
• Claude Haiku 5.5 migration guide
• Anthropic API pricing
• Claude Haiku 5.5 system card

For the current small-model pair, see our Haiku 5.5 vs Luna 6 comparison for pricing thresholds, API fit and costs on the same token workload.