Skip to content
AIMODELRACE
AnthropicLanguage modelNew

Claude Haiku 5.5

Tiny price, startling computer-use skill: the sub-agent and high-volume specialist.

Compare head-to-headRace it in 3D

Overall Intelligence

#23

of 56
9.2 pts behind #1

1

races led

2

weeks at #1

79.5

score

Released
Oct 7, 2026
2 wk ago
Context
1M
128K max output
Price
$0.1 / $0.5
per 1M in/out · $0.20 blended
Speed
262 tok/s
0.7s to first token
Modalities
text · image
outputs text
Weights
Closed
API and apps

The 30-second version

Anthropic's smallest model, priced like a commodity but with surprisingly strong computer-use skills. Haiku 5.5 is the go-to sub-agent and high-volume specialist.

Key takeaways
  • 1Costs $0.10 / $0.50 per million tokens: cheap enough to run on every request.
  • 2Remarkably good at operating software through screenshots and clicks for its size.
  • 3Over 250 tokens per second: ideal for real-time features.
  • 4Not the model for deep reasoning or long, ambiguous tasks.

What you need to know

Purpose

Fast, cheap intelligence for classification, extraction, routing, sub-agent work and computer-use automation at scale.

What Anthropic optimised for

Latency and cost, with targeted investment in tool use and computer use so it can serve as a capable worker inside larger agent systems.

Unique strengths

  • Computer use well above its weight class.
  • Very low latency.
  • Reliable structured output for extraction and routing.
Insight
How it differs from rivals. Versus GPT-6 Luna and Gemini 3.5 Flash it is cheaper and stronger at computer use; Flash is faster and better at multimodal input, Luna is the friendlier chat model.

Where it ranks

Current snapshot · illustrative scores · click any race to watch it

CodingHard ReasoningMathScienceWriting & LanguageMultimodalTool Use & AgentsLong ContextTrust & FactualitySpeed & Cost

Haiku 5.5Best on each axis

Overall Intelligence over time

#1 frontierHaiku 5.5Qwen3.6-MaxGPT-6 Luna
55.060.065.070.075.080.085.090.02025Apr ’25Jul ’25Oct ’252026Apr ’26Jul ’26Oct ’26

Best use cases

01

Sub-agents

Parallel workers that search, read and summarise for a larger orchestrator.

02

Browser and desktop automation

Form filling, data entry and UI testing driven by screenshots.

03

Classification and extraction

Tagging tickets, pulling fields from documents and routing requests.

Strengths

  • Price

    Among the cheapest capable models available.

  • Computer use

    Competitive with much larger models on screen-driven tasks.

  • Latency

    Fast first token and high throughput.

Weaknesses

  • Reasoning depth

    Falls well behind on hard math, science and frontier reasoning.

  • Long tasks

    Loses the thread on long, multi-stage work without an orchestrator.

Prompts that play to its strengths

Ticket triage

Classify each support ticket into exactly one of: billing, bug, feature_request, account, other. Return JSON lines: {"id": ..., "label": ..., "confidence": 0-1}. If confidence < 0.6, use "other".

<tickets>...</tickets>

Why it works: Narrow, structured tasks with a strict format are its sweet spot.

UI regression check

Open the staging site, log in with the test account, add any item to the cart and complete checkout with the test card. Report each step with a screenshot description and flag anything that looks broken or slow.

Why it works: Uses its standout computer-use ability on a well-defined flow.

When to choose it, and when not to

Haiku 5.5overGPT-6 Luna

You need computer use or the lowest price per token.

Haiku 5.5over3.5 Flash

The task is screen automation rather than raw speed or video input.

Pick something else when…

  • Complex reasoning or writing that needs nuance.
  • Tasks that require sustained planning over many steps without supervision.

Update history

The distilled version first; open any release for the deeper breakdown.

The Claude Haiku family

Haiku 4.5October 2025Haiku 5.5October 2026