Claude Haiku 5.5 Tops Small Models, Beats GPT-6 Luna at 90% Lower Cost

Introduction

Anthropic released Claude Haiku 5.5, a small model that now benchmarks above DeepSeek V4.1 Flash and GLM-5.3-Flash, and goes head-to-head with GPT-6 Luna across multiple tests. The model targets high-frequency, low-latency, cost-sensitive tasks, complementing Opus 5.5 (complex reasoning) and Sonnet 5.5 (general execution).

Performance Benchmarks

Official comparisons show Haiku 5.5 outperforming DeepSeek V4.1 Flash and GLM-5.3-Flash across multiple benchmarks, and matching or beating GPT-6 Luna item by item.

Pricing and Tokenizer Impact

For requests under 100k tokens, input and output prices drop 90% versus Haiku 4.5; above 100k tokens, a 50% reduction. However, Haiku 5.5 adopts the same tokenizer as Sonnet 5.5 and Opus 5.5, which can increase token counts for code, tables, and non-English text, partially offsetting the per-token savings. The article recommends measuring actual task cost with four metrics: input tokens, output tokens, success rate, and total cost per task.

Five-Tier Effort Configuration

Haiku 5.5 introduces effort settings from low to max, allowing developers to tune reasoning intensity per task without switching models. This reduces context-switching overhead in agent workflows.

Computer-Use Performance (OSWorld 2.1)

Haiku 5.5 Low: 42.0% accuracy, $0.07 per task

Haiku 5.5 Max: 72.4% accuracy, $0.61 per task

Haiku 4.5 Max: 15.7% accuracy, $1.45 per task

Even the low setting surpasses the previous generation's max, at half the cost.

Professional Task Performance (GDPval-AA v2.1)

Haiku 5.5 Low: Elo 1125, $0.01 per task

Haiku 5.5 Max: Elo 1620, $0.87 per task

Haiku 4.5 Max: Elo 735, $0.24 per task

The trend confirms that low effort handles lightweight tasks cheaply, while high effort tackles complex execution.

Limitations vs Sonnet 5.5

Terminal-Bench 4.0 shows Haiku 5.5 at 39.2% versus Sonnet 5.5 at 70.6% for complex multi-step coding, cross-file refactoring, and long-horizon autonomous planning. Anthropic recommends Sonnet 5.5 or Opus 5.5 for such tasks. Haiku 5.5 fits the execution layer: parallel test runs, code checks, log classification, and verified sub-tasks. Cognition's Devin combines Opus 5.5 as primary with Haiku 5.5 as sub-agents, achieving 66.2% on FrontierCode, higher than either alone.

Migration from Haiku 4.5: Five API Changes

budget_tokens

replaced by adaptive thinking and effort parameter; old parameter causes errors.

Sampling parameters temperature, top_p, top_k locked to defaults; applications relying on them must redesign generation logic.

Assistant message prefill removed; must use tool calls or structured output interfaces for format control.

Computer tool updated from computer_20250124 to computer_toolset_20260801; interface and response structure may change.

Adaptive thinking enabled by default; first response block may be a thinking block, not final text. Parsing must filter by type field.

Additional Anthropic Updates

Sonnet 5.5 cache read price cut from $0.20 to $0.10 per million tokens, reducing typical agent costs ~20%. Max and Team subscribers receive monthly API credits: Max 5x $100, Max 20x $200, Team up to $500 shared.

Recommended AI Coding Workflow

Use Haiku 5.5 as execution-layer sub-agent: primary model handles requirements, planning, and complex errors; Haiku 5.5 runs parallel tests, checks code, classifies failures, generates glue code, and verifies sub-task results. Do not assign cross-file refactoring or long-horizon planning to Haiku 5.5.

FAQ Highlights

Haiku 5.5 cannot fully replace Sonnet 5.5.

Price cut 90% for ≤100k tokens, 50% above; actual cost depends on tokenizer-induced token inflation.

Five effort tiers from low to max.

Migration requires code changes for all five API differences.

Suitable for Claude Code as sub-agent for parallel execution, not primary planning.