← Back to blog
Claude Haiku 5.5 Review: Price, Benchmarks & When to Use It

Claude Haiku 5.5 Review: Price, Benchmarks & When to Use It

Alex Morgan

By Alex Morgan

MyClaw Editorial

Claude Haiku 5.5

Try Claude Haiku 5.5

Start a chat with Claude Haiku 5.5 for coding, writing, or research.

Chat with Claude Haiku 5.5

AI Takeaway

  • What is Claude Haiku 5.5 and when did it come out? Claude Haiku 5.5 is the small, fast member of Anthropic's Claude 5.5 family, released on October 7, 2026. Its model ID is claude-haiku-5-5. It reads up to 1M tokens, writes up to 128K, and is the first Haiku with an effort setting.
  • How much does Haiku 5.5 cost? Anthropic charges $0.10 for each million input tokens and $0.50 for each million output tokens on prompts up to 100K tokens, and $0.50 / $2.50 above that. Batch requests get 50% off. A newer tokenizer also turns identical text into roughly 30% more tokens.
  • How does Haiku 5.5 score on benchmarks? Per Anthropic, it reaches 72.4% on OSWorld 2.1 (Haiku 4.5: 15.7%), 39.2% on Terminal-Bench 4.0 (Haiku 4.5: 0.0%) and 1620 Elo on GDPval-AA (Haiku 4.5: 735). Sonnet 5.5 still scores higher on every published row.
  • Should you switch from Haiku 4.5 to Haiku 5.5? For most workloads, yes. Anthropic's list prices fall by 90% below 100K tokens, and the scores jump. Recount your prompts and fix the breaking API changes before you switch.
  • Can you run Claude Haiku 5.5 on MyClaw? Yes. A hosted OpenClaw or Hermes Agent agent on MyClaw can use Haiku 5.5 for its high-volume steps while Sonnet 5.5 takes the hard ones.

Claude Haiku 5.5 Specs, Model ID and Availability

Anthropic built Haiku 5.5 for work where volume and cost matter more than peak reasoning. It is the third Claude 5.5 model.

SpecClaude Haiku 5.5Claude Haiku 4.5
ReleasedOctober 7, 2026October 15, 2025
Model IDclaude-haiku-5-5 (Bedrock: anthropic.claude-haiku-5-5)claude-haiku-4-5
Context / max output (tokens)1M / 128K200K / 64K
ThinkingAdaptive, effort from low to max (default medium)Extended thinking with a token budget
Price per 1M tokens (Anthropic)From $0.10 input / $0.50 output$1 input / $5 output

A context window five times larger means one request can carry a large slice of a codebase. Effort control lets you turn it down for quick classification and up for multi-step tool use.

You can pick Haiku 5.5 in Claude.ai on every plan, Free through Enterprise. Developers get it on Anthropic's API, AWS, Google Cloud and Microsoft Foundry. It landed nine days after Claude Sonnet 5.5 and matches that model's 1M context and 128K output.

Introducing Claude Haiku 5.5 \ Anthropic

Haiku 5.5 Price: Two Tiers Split at 100K Tokens

Anthropic's list prices per 1M tokens, with a break at 100K prompt tokens:

Item (Anthropic)Haiku 5.5, prompts up to 100KHaiku 5.5, prompts over 100KHaiku 4.5
Input$0.10$0.50$1.00
Output$0.50$2.50$5.00
Cache read$0.01$0.05$0.10
5-minute cache write$0.125$0.625$1.25
Batch input / output$0.05 / $0.25$0.25 / $1.25$0.50 / $2.50

On average, Anthropic says, Haiku 5.5 costs about 75% less to run than Haiku 4.5. Rates drop 90% for requests up to 100K tokens and 50% above it, and roughly nine in ten Haiku 4.5 requests fit under that line. Sonnet 5.5's input rate, $2 per 1M tokens on Anthropic's price sheet, is twenty times the short-prompt Haiku rate.

Why the New Tokenizer Raises Real Costs

Haiku 5.5 splits text into more pieces than Haiku 4.5 did. Anthropic puts the increase at about 30% for the same text, varying by content. That pushes some prompts over the line: a 90K-token Haiku 4.5 prompt recounts to about 117K (computed) and lands in the higher tier.

Computed example from Anthropic's list prices: a job with 50K input and 5K output tokens costs $0.075 on Haiku 4.5. Add 30% on both sides and the same job becomes about 65K in and 6.5K out on Haiku 5.5, or roughly $0.01. The saving holds only while the prompt stays under 100K, so recount real prompts and use caching and batch for repeat work. Sonnet 5 made the same tokenizer switch last cycle.

Haiku 5.5 Benchmarks: A Big Jump Over Haiku 4.5

Every score below is per Anthropic. None has been reproduced independently yet, and Anthropic doesn't say which effort level produced them.

Benchmark (per Anthropic)Haiku 5.5Haiku 4.5GPT-6 LunaSonnet 5.5
GDPval-AA v2.1 (Elo)162073514371840
AA-Briefcase v1.1 (Elo)157861413361824
OSWorld 2.1, offline subset72.4%15.7%48.9%83.9%
Humanity's Last Exam (no tools)45.9%10.2%—56.9%
Humanity's Last Exam (tools)57.4%18.7%—64.5%
Terminal-Bench 4.039.2%0.0%16.4%70.6%
FrontierCode 1.1 (Main)46.4%—42.4%52.1% (Xhigh)
Chartography, no tools46.4%6.4%29.1%61.6%

Computer use shows the biggest swing, a 56.7-point gain on OSWorld 2.1. Agentic coding is close behind: Haiku 4.5 scored zero on Terminal-Bench 4.0, and Haiku 5.5 clears 39%. OpenAI's high-volume model trails on every row where both have a score; our breakdown of GPT-6 Luna shows where it sits in that lineup. AlphaSense ran 400 queries through its document Q&A feature and reports 0.84 for Haiku 5.5 against 0.76 for Haiku 4.5.

Claude Haiku 5.5 (max) - Intelligence, Performance & Price Analysis |  Artificial Analysis

Where the Bigger 5.5 Models Still Win

Sonnet 5.5 comes closest on FrontierCode (46.4% vs 52.1%) and pulls furthest ahead on Terminal-Bench 4.0 (39.2% vs 70.6%). Anthropic still steers complex agentic coding to its larger models. Its system card says Haiku 5.5 rivaled or exceeded Sonnet 5.5 only "in a few areas." For long coding runs, see how Opus 5.5 and Sonnet 5.5 divide that work.

Upgrading From Haiku 4.5: Six API Changes

Anthropic's migration guide flags these changes beyond the model ID:

  • Thinking: a fixed budget_tokens value now fails with a 400 error. Switch to adaptive thinking, and pick a low effort where Haiku 4.5 ran without thinking.
  • Sampling: top_k, top_p and temperature should be left out; values other than the defaults return a 400 error.
  • Prefill: a final assistant turn is rejected. Use structured outputs, or move that text into the user turn.
  • Computer use: on Anthropic's API and Google Cloud, a new toolset replaces the old computer tool.
  • Limits: thinking counts toward max_tokens, so a small cap can end a reply before any text appears.
  • Refusals and capacity: handle the new "refusal" stop reason, and plan without Priority Tier, which Haiku 5.5 doesn't support.
Put Claude Haiku 5.5 to work on a real task

Start a chat with Claude Haiku 5.5 for coding, writing, or research.

Chat with Claude Haiku 5.5

Best Uses for Claude Haiku 5.5

Anthropic recommends Haiku 5.5 when speed and volume matter most. That covers:

  • High-volume text work: classification, extraction, routing and summaries.
  • Live chat and support: latency-sensitive replies, voice agents and in-app assistants.
  • Subagents: a larger model plans while Haiku 5.5 runs scoped pieces in parallel, including multi-step tool use.
  • Browser and desktop chores: form filling, data entry and moving records between apps.
  • Small code edits: direct, well-defined changes.

Long autonomous coding sessions are a poor fit, as are prompts that routinely pass 100K tokens. Choosing a default model for an agent? Our notes on picking a model for OpenClaw agents cover how small and large models share a workload.

Claude Haiku 5.5 Review Verdict

Haiku 5.5 is the best small Claude by a wide margin and the default pick for volume work. At Anthropic's list prices it costs a tenth of Haiku 4.5 per token on short prompts, and it now handles agent tasks where Haiku 4.5 scored near zero. Keep a larger model in charge of hard agent jobs.

Pros:

  • Lowest per-token price in Anthropic's current lineup below 100K
  • Effort control from low to max
  • Big gains over Haiku 4.5 on every reported benchmark
  • Fastest Claude at standard speed, per Anthropic

Cons:

  • Price jumps 5x above 100K tokens
  • New tokenizer inflates token counts
  • Several breaking API changes
  • Trails Sonnet 5.5 on every reported benchmark
  • Scores are self-reported so far

Put Claude Haiku 5.5 on an Agent's Routine Work

MyClaw hosts OpenClaw and Hermes Agent, and both can run Claude Haiku 5.5. Give Haiku 5.5 the frequent, scoped steps, such as summaries, triage, lookups and routing, and leave planning and hard reasoning to Sonnet 5.5. Running Hermes Agent? Our Hermes Agent model picks show where each model fits.

Get Started

Three Steps to Start on MyClaw

  1. Start a hosted agent on OpenClaw or Hermes Agent.
  2. Choose Haiku 5.5 as the agent's main model, connect your chat app, and keep Sonnet 5.5 for the hard steps.
  3. Hand it one recurring task, then review output, retries and token use after a week.

Put Claude Haiku 5.5 to work on a real task

Start a chat with Claude Haiku 5.5 for coding, writing, or research.

Chat with Claude Haiku 5.5
Claude Haiku 5.5 Review: Price, Benchmarks & When to Use It | MyClaw.ai