Last updated: October 8, 2026

Claude Haiku 5.5: benchmarks, price and how to use it

Anthropic released Claude Haiku 5.5 on October 7, 2026. It is the small, fast model of the Claude 5.5 family, priced at $0.10 per 1M input tokens. Here is what it is good at, what the benchmarks say, what it costs in practice, and when Sonnet 5.5 or Opus 5.5 is the better pick.

On this page: Claude Haiku 5.5 benchmark results · Claude Haiku 5.5 pricing · How to use Claude Haiku 5.5 · Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5

Key facts

Source: Anthropic: Claude models overview

Claude Haiku 5.5 benchmark results

Anthropic's launch table. Sonnet 5.5 is included for reference, since it is the next model up.

• Haiku 5.5's 39.2% on Terminal-Bench 4.0 is at max effort. At the default Medium effort it scores about 20% (VentureBeat).
• OSWorld 2.1 here is the offline subset, so it is not comparable with the full OSWorld 2.1 numbers on our Sonnet 5.5 and Opus 5.5 pages.
• Sonnet 5.5's FrontierCode score is at Xhigh effort. "—" means Anthropic did not report that score.
• These are Anthropic's numbers and have not been independently reproduced. GDPval-AA and AA-Briefcase are Elo ratings, higher is better.

Source: Anthropic: Claude Haiku 5.5 launch page

Independent score

Artificial Analysis gives Haiku 5.5 (max effort) 43 on its Intelligence Index, rank 43 of 696 models, against a median of 12 for models in its class. It also used 440M output tokens to run the index, against a median of 87M: at max effort the model thinks a lot, so the low token price does not automatically mean a low bill.

Source: Artificial Analysis: Claude Haiku 5.5

What launch customers reported

• HubSpot: 92.8% averaged over three runs on its CRM test suite, the best score it has seen there.
• AlphaSense: 0.84 against 0.76 for Haiku 4.5 over 400 queries in its Ask in Document feature, which handles about 8M calls a week.
• Box: 11 points above Haiku 4.5 at about half the latency.
• Asana: over 30% lower task-completion latency and up to 2.5x faster inference per agent turn than its current model.
• Cognition: with Haiku 5.5 as the helper model under Opus 5.5, Devin Fusion scored 66.2 on FrontierCode at lower cost and latency.

Source: Anthropic: Claude Haiku 5.5 launch page

Claude Haiku 5.5 pricing

USD per 1M tokens from Anthropic's pricing page. Haiku 5.5 is the only current Claude model priced by prompt length: a prompt over 100,000 tokens pays the higher row for the whole request.

• Batch API: 50% off input and output.
• Against Haiku 4.5 ($1.00 / $5.00) that is 90% cheaper up to 100K tokens and 50% cheaper above. Anthropic estimates about 75% lower cost on average, because about 90% of Haiku 4.5 requests were under 100K tokens.
• The new tokenizer counts the same text as roughly 30% more tokens than Haiku 4.5 (MarkTechPost), so re-measure before you compare bills.

Source: Anthropic: Claude API pricing

Price next to similar models

What one task really costs

Same assumptions as our calculator (60% of input from cache), one agent task by codebase size. The two Haiku 5.5 columns show the price if each request stays under 100K tokens and if it goes over.

At 173 tasks a month (8 a day, 5 days a week, medium codebase) that is about $2.23 a month on Haiku 5.5 with short prompts, against $43.08 on Sonnet 5.5.

Compare every model in the cost calculator

How to use Claude Haiku 5.5

• In the Claude apps: pick Haiku 5.5 in the model menu. It is available on every plan, Free included.
• In Claude Code: switch the model to Haiku 5.5, or let a larger model hand it subtasks as a subagent.
• On the API: use the model ID claude-haiku-5-5 and set the effort level per request. Low or Medium suits classification, extraction and chat; raise it for multi-step tool use.
• On cloud platforms: Amazon Bedrock, Google Cloud and Microsoft Foundry list it at launch.

Python example with the official Anthropic SDK:

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

msg = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4096,
    output_config={"effort": "low"},  # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Classify this ticket: ..."}],
)
print(msg.content[-1].text)

Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5

Anthropic positions Haiku 5.5 as the fast, cheap model for scoped work and as a subagent under the larger models.

Moving from Haiku 4.5

• Change the model ID to claude-haiku-5-5.
• Remove non-default temperature, top_p and top_k values: they now return a 400 error.
• Thinking budgets (budget_tokens) are gone. Thinking is adaptive and controlled through the effort level.
• Re-measure token counts: the tokenizer is new. Keep an eye on prompt length so requests stay in the cheaper tier under 100K tokens.
• Cyber safeguards are stricter than on Haiku 4.5: penetration testing and attacker-style requests are blocked under standard access.

Claude Sonnet 5.5 · Claude Opus 5.5 · GPT-6 Sol / Luna · All model prices · Cost calculator

Independent page. Not affiliated with Anthropic. Benchmark figures are Anthropic's or the named source's; we have not run our own tests.