Claude Haiku 5.5 in 1 Page — 5W1H Breakdown & Benchmark Scorecard

Released on October 7, 2026, Claude Haiku 5.5 (claude-haiku-5-5) brings the full Claude 5.5 architecture—1M context, 128K output, and adaptive thinking—to the bottom of the price ladder at $0.10 /$0.50 per 1M tokens (for prompts ≤100K). Here is the complete 1-page 5W1H engineering and benchmark brief.

1. Executive Summary: Claude Haiku 5.5 in 5W1H

2. WHY: Official & Real-World Benchmark Scorecard

Haiku 5.5 transitions from a text-only lightweight classifier into a capable agentic worker—outperforming Haiku 4.5 by massive margins and leading OpenAI GPT-6 Luna across every shared evaluation, while trailing Sonnet 5.5 on complex coding.

Head-to-Head Model Benchmarks

Real-World Production & Hands-On Benchmarks

• DataCamp 24-Invoice Accounts Payable Test: Tested on 24 vendor invoices with policy traps (3% price tolerance, split deliveries, swapped PO digits). Both Haiku 5.5 (high effort) and Haiku 4.5 (16K thinking budget) scored 24/24 (100% accuracy), but Haiku 5.5 used under 25% of the output tokens and cost $0.0098 vs. $0.25 (~25x cheaper).
• Enterprise Production Evaluations:
  • HubSpot (CRM Audit): 92.8% average score across 3 runs—fastest completion, highest hit rate, and lowest false-positive rate.
  • AlphaSense (8M weekly doc queries): Statistically significant jump from 0.76 → 0.84 over Haiku 4.5.
  • Box AI & Asana: +11 points higher accuracy at ~50% latency (Box); 20% lower task latency and up to 2.5x faster inference per agent turn (Asana).
  • Cognition (Devin Fusion): Paired as a sidekick to Opus 5.5, holds a 66.2 FrontierCode score while slashing cost and latency.

Pricing & Token Economics (Per 1M Tokens)

Why "75% cheaper on average" instead of 90%? Two reasons: (1) Haiku 5.5 uses the newer 5.5-family tokenizer, which produces ~30% more tokens for the same text than Haiku 4.5, and (2) prompts over 100K tokens (~10% of historical traffic) pay 5x the base rate ($0.50/$2.50). Combined with the 50% Batch API discount, high-volume pipelines still see dramatic savings.

3. WHO, WHERE & WHEN: Claude 5.5 Tier & Effort Matrix

All four current Claude models share a 1M context window and 128K max output:

4. HOW: Minimal Implementation & Breaking Migration Checklist

from anthropic import Anthropic

client = Anthropic()
response = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4096,
    output_config={"effort": "low"},  # "low" | "medium" (default) | "high" | "max"
    messages=[{"role": "user", "content": "Extract invoice total and PO number: ..."}],
)
# Adaptive thinking is on by default; filter blocks by type rather than index 0
print(next(b.text for b in response.content if b.type == "text"))

5 Breaking Changes When Migrating from Haiku 4.5

1. No budget_tokens: Manual extended thinking throws an error; switch to output_config={"effort": ...}.
2. No Custom Sampling: Non-default temperature, top_p, or top_k values return an API error.
3. No Assistant Prefill: Prefilling assistant messages is no longer supported.
4. Read Content by Type: Responses may start with thinking blocks—always select b.type == "text" instead of response.content[0].
5. Updated Computer Use Toolset: Requires computer_toolset_20260801 and append-only conversation history (editing prior turns invalidates thinking blocks).

References

• Anthropic — Introducing Claude Haiku 5.5
• DataCamp — Claude Haiku 5.5: Features, Benchmarks, Pricing, and Hands-On Test