Sign in Subscribe
news

Claude Haiku 5.5

Sebastien Dubois

Sebastien Dubois

— 7 min read

Canonical version: Claude Haiku 5.5.

Claude Haiku 5.5 is Anthropic's small, fast Claude model, released on 7 October 2026. It's the third model of the Claude 5.5 family, after Claude Opus 5.5 (22 September) and Claude Sonnet 5.5 (28 September), and it replaces Claude Haiku 4.5, which was almost a year old. Model ID: claude-haiku-5-5 (anthropic.claude-haiku-5-5 on Amazon Bedrock). It has no date suffix and no separate alias. See Claude 5 for how the tiers fit together.

Anthropic built it for high-volume, latency-sensitive work: classification, routing, extraction, summaries, compaction and subagent tasks. It's the fastest model in the lineup at standard speed (Opus in fast mode is quicker). Anthropic's own advice is to pair it with Opus 5.5 or Sonnet 5.5 as a subagent on coding work, and to leave complex agentic coding to the bigger models.

For prompts up to 100K tokens, it costs a tenth of Haiku 4.5. Above 100K, every token of the request costs five times more. That threshold is the thing to design around.

Pricing

Per million tokens:

Prompt ≤ 100K tokens Prompt > 100K tokens Haiku 4.5
Input $0.10 $0.50 $1
Output $0.50 $2.50 $5
Cache reads $0.01 $0.05 $0.10
Cache writes (5 min) $0.125 $0.625 $1.25
Cache writes (1 h) $0.20 $1 $2
  • The Batch API halves these prices
  • Minimum cacheable prompt: 512 tokens, down from 4,096 on Haiku 4.5
  • Each request is priced on its own, and the prompt length counts every input token, cache reads and cache writes included. A 120K-token prompt that is 90% cache hits still pays the higher rates on all its tokens, output included
  • Anthropic says about 90% of Haiku 4.5 requests stayed under 100K tokens, and that Haiku 5.5 costs around 75% less to run on average once the new tokenizer is counted

Under 100K tokens, that's exactly the price of GPT-6 Luna ($0.10 / $0.50). Above it, Luna is much cheaper: its price only goes up past 272K tokens, and only to $0.20 / $0.75.

The tokenizer

Haiku 5.5 uses the tokenizer introduced with Claude Opus 4.7. According to Anthropic, the same text produces about 30% more tokens than on Haiku 4.5; the exact increase depends on the content. Simon Willison measured about 1.25x on a long prompt and calls it a "hidden price increase". It also means you reach the 100K threshold sooner than your Haiku 4.5 token counts suggest, so recount prompts with the new model before estimating costs (see AI Tokenization).

Images cost more too: Haiku 5.5 uses the high-resolution image tier, and a 2000×1500 image takes about 2.5x the visual tokens it did on Haiku 4.5.

Specs

  • Context window: 1M tokens (Haiku 4.5: 200K)
  • Max output: 128K tokens (Haiku 4.5: 64K), or 300K on the Message Batches API with the output-300k-2026-03-24 beta header
  • Input: text and images. Output: text
  • Knowledge and training cutoff: June 2026
  • Platforms: Claude API, Amazon Bedrock, Google Cloud (Vertex AI), Microsoft Foundry and Claude Platform on AWS
  • Retirement: not before 7 October 2027
  • No Priority Tier, and no structured outputs on Amazon Bedrock

Thinking and effort

It's the first Haiku with an effort setting. Adaptive thinking is on by default: the model decides when and how much to think. The five levels are low, medium (the default), high, xhigh and max. Anthropic recommends medium for most work, agentic coding included, and low for chat, short tool tasks and simple high-volume requests. At low, it can skip thinking entirely on simple requests. See Claude Code effort levels.

Simon Willison writes that it "doesn't let you disable reasoning". Anthropic's "What's new" page says thinking: {"type": "disabled"} still works at high effort or below, and recommends lowering effort instead.

Benchmarks

Anthropic-published, so read them as directional:

  • Terminal-Bench 4.0 (agentic coding): 39.2%, against 0% for Haiku 4.5, 16.4% for GPT-6 Luna and 70.6% for Sonnet 5.5
  • FrontierCode 1.1 Main: 46.4% (Luna: 42.4%, Sonnet 5.5 at Xhigh: 52.1%)
  • OSWorld 2.1, offline subset (computer use): 72.4% (Haiku 4.5: 15.7%, Luna: 48.9%)
  • GDPval-AA v2.1 (knowledge work, Elo): 1620 (Haiku 4.5: 735, Luna: 1437, Sonnet 5.5: 1840)
  • Humanity's Last Exam with tools: 57.4% (Haiku 4.5: 18.7%, Sonnet 5.5: 64.5%)

It beats Luna everywhere Anthropic compared them. The Terminal-Bench gap with Sonnet 5.5 (39.2% vs 70.6%) is why Anthropic keeps complex agentic coding on Sonnet and Opus.

The customer quotes describe the same split. At Rogo, a bigger model builds the deck while a Haiku 5.5 subagent pulls the segment revenue line out of a 10-K. Cognition uses it as the "sidekick" next to Opus 5.5 in Devin. Asana measured over 30% lower latency on its agent tasks, and AlphaSense scored it 0.84 vs 0.76 for Haiku 4.5 on 400 document questions.

Migrating from Haiku 4.5

The model ID swap alone breaks code:

  • Thinking with budget_tokens returns a 400. Use adaptive thinking and set the effort level
  • Omit temperature, top_p and top_k. Only temperature: 1 and top_p: 0.99 are accepted; any other value returns a 400
  • Assistant prefill returns an error: end messages with a user turn
  • Responses can start with thinking blocks, so select content blocks by type instead of reading content[0]
  • Thinking tokens count toward max_tokens. A small limit can stop after the thinking block, before any text
  • Thinking blocks from earlier turns stay in context and count as input tokens. Together with the tokenizer, that makes multi-turn conversations grow faster, toward the 100K threshold
  • Thinking blocks only work in the account that produced them, or in a linked account
  • Safety classifiers can decline a request (stop_reason: "refusal"), and there's no server-side fallback
  • Computer use moves from computer_20250124 to the computer_toolset_20260801 toolset on the Claude API and Google Cloud

Safeguards

Anthropic reports far fewer misaligned behaviors than Haiku 4.5 and a lower willingness to cooperate with misuse. Its cybersecurity safeguards are stricter than Haiku 4.5's but looser than those of the other recent models: defensive work gets more room than on Sonnet 5.5, while penetration testing stays blocked. Biology safeguards are the same as on Sonnet 5, Sonnet 5.5 and Opus 5.

The pelican test

Simon Willison ran his pelican-on-a-bicycle SVG test through llm and llm-anthropic 0.30, the first plugin release that picks up new models without a new version (llm anthropic refresh). At low effort, it took 7 seconds and cost 0.0936 cents, with a weak bicycle frame. Every level above low drew a good frame. At max, it took 5 minutes 9 seconds and cost 3.38 cents, and the reasoning trace opened with "This is the classic pelican-on-bicycle SVG test...", so the model knows the benchmark. Haiku 4.5's pelican, a year earlier, cost 0.76 cents and, in his words, "sucked".

On the same day, Anthropic halved Sonnet 5.5's cache reads and added monthly API credits for Max and Team subscribers (see Claude).

My take

This is the model for the cheap steps of a pipeline: classifiers, routers, extractors and subagents that do one well-defined thing (Model routing, AI Subagents). Haiku 4.5 was hard to justify at $1 / $5 when Luna cost a tenth of that. Haiku 5.5 closes that gap, as long as each request stays under 100K tokens.

It matters to me because 247 of the 448 skills in my vault declare model: haiku in their frontmatter. In Claude Code v2.1.293 and later, the haiku alias resolves to Haiku 5.5 on the Anthropic API, so those skills switch models without any edit. On Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, the alias still points to Haiku 4.5.

In May 2026, Haiku 4.5 skills that loaded two or more heavy context loaders crashed with "Context limit reached", and I moved 24 of them to Sonnet. With a 1M window, those skills could go back to Haiku. If they overflowed 200K tokens, though, they will pay the over-100K rates: $0.50 / $2.50, still four times cheaper than Sonnet 5.5.

How I'd design around the 100K threshold:

  • Give each subagent one task and only the files it needs, not the parent's whole conversation
  • Prefer fresh subagents over one long-lived Haiku conversation. Old thinking blocks pile up in context, and a session crosses 100K without you noticing (thinking block clearing helps)
  • Don't count on caching to stay under the line: cache hits count toward the threshold
  • Track prompt sizes per request in your logs, recounted with the new tokenizer
  • For single requests over 100K (a long document, a big codebase dump), compare with Luna. It's cheaper there. Mercury 3 is another candidate: it keeps the same price per token up to 1M

Before replacing Haiku 4.5 in a classifier or a judge, run both side by side on real traffic first (Shadow Evaluation). Better benchmark scores don't guarantee the same decisions on your data.

Caveats

  • Benchmark figures and customer quotes come from Anthropic's announcement. The only independent numbers here are Simon Willison's pelican runs and token count
  • Anthropic's docs and Simon Willison disagree on whether thinking can be disabled (see above)
  • The 30% token increase is Anthropic's average; Simon measured 1.25x on one prompt
  • In Claude Code on third-party clouds, haiku still means Haiku 4.5

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe