Claude Haiku 4.5

Anthropic · claude family · Official Docs

Claude Haiku 4.5 is the cost-efficiency champion of the Claude family — at $1/$5 per MTok, it delivers surprisingly frontier-adjacent quality for high-volume workloads. The model's biggest gotcha is verbosity: it generates roughly 2x the tokens of comparable models on equivalent tasks, which can erode cost savings if left unchecked. Refrase's adaptation layer addresses this by injecting conciseness directives and output-format constraints that keep responses focused. The XML structuring gains that benefit Opus and Sonnet apply equally to Haiku, with our eval showing a ~10% improvement from prompt adaptation. Haiku 4.5 is the first in its class to support extended thinking and computer use, making it viable for agentic workflows that were previously Sonnet-only territory. For judge/evaluator roles in multi-model pipelines, Haiku 4.5 offers excellent cost-performance with zero timeouts in our 46-model eval study (replacing DeepSeek V3.2 which had ~17% timeout rate). The 200K context limit is its main constraint — for document-heavy extraction exceeding that window, Sonnet 4.6 at 3x the price is the next step up.

#15

Rank

Quality Score

450ms

Avg Response

+10%

Adaptation Gain

Specifications

200K

Context Window

64K

Max Output

$1 / $5

Per 1M tokens (in/out)

Batch API: $0.50/$2.50 per MTok (50% discount). Prompt caching: 5-min write 1.25x, 1-hour write 2x, cache hit 0.1x base input. Extended thinking tokens billed as output at $5/MTok. Does not support adaptive thinking. (source: Anthropic docs, Pricing; Models Overview)

Key Capabilities

✓Fastest model with near-frontier intelligence — achieves roughly 90% of Sonnet 4.5's capabilities at a fraction of the cost (source: Anthropic docs, Models Overview; DataCamp, Claude Haiku 4.5)
✓First Haiku model to include extended thinking, computer use, and context awareness (source: Anthropic, Introducing Claude Haiku 4.5)
✓73.3% on SWE-bench Verified — highest after Sonnet 4.5, surpassing even Sonnet 4 (source: Anthropic, Introducing Claude Haiku 4.5)
✓50.7% success rate on computer use benchmarks, outperforming Sonnet 4's 42.2% (source: Anthropic, Introducing Claude Haiku 4.5)
✓Structured outputs with guaranteed JSON schema compliance via json_schema format and strict tool use (source: Anthropic docs, Structured Outputs)
✓200K context window with 64K max output tokens (source: Anthropic docs, Models Overview)

Known Limitations

⚠Does not support adaptive thinking — only manual extended thinking with budget_tokens (source: Anthropic docs, Models Overview)
⚠200K context window only — no 1M token option available unlike Sonnet 4.6 and Opus 4.6 (source: Anthropic docs, Models Overview)
⚠Notably verbose output — generates approximately 2x the tokens compared to average models on equivalent benchmarks, which increases effective output costs (source: DataCamp, Claude Haiku 4.5; Artificial Analysis)
⚠Small but meaningful gaps compared to frontier models on multi-hop reasoning and highly nuanced analysis tasks (source: DataCamp, Claude Haiku 4.5)
⚠Middle-of-pack on LiveCodeBench and academic reasoning benchmarks despite strong agentic coding performance (source: DataCamp, Claude Haiku 4.5)

Prompt Patterns

Preferred Instruction Format

XML tags (<instructions>, <context>, <output_format>, <examples>) for structured prompts. System prompt via the 'system' API parameter. Role-setting in system prompt focuses behavior and tone.

Recommended Practices

Use XML tags to structure complex prompts — same best practices as Sonnet and Opus apply (source: Anthropic docs, Prompting Best Practices, Structure Prompts with XML Tags)
Use manual extended thinking with budget_tokens for complex tasks that benefit from step-by-step reasoning; minimum budget is 1,024 tokens (source: Anthropic docs, Extended Thinking)
Provide 3-5 diverse examples in <example> tags for few-shot prompting (source: Anthropic docs, Prompting Best Practices, Use Examples Effectively)
Add explicit conciseness instructions to counteract verbosity — e.g., 'Respond concisely. Avoid unnecessary elaboration.' (source: Refrase eval, verbosity analysis across Haiku 4.5 outputs)
Place longform data at the top of prompts, above queries and instructions (source: Anthropic docs, Prompting Best Practices, Long Context Prompting)

Anti-Patterns to Avoid

Do not use adaptive thinking — Haiku 4.5 does not support it; use manual thinking with budget_tokens instead (source: Anthropic docs, Models Overview)
Avoid exceeding 200K input tokens — Haiku has no 1M context option and will reject oversized requests (source: Anthropic docs, Models Overview)
Avoid relying on Haiku for multi-hop reasoning or highly nuanced analysis where frontier accuracy is required (source: DataCamp, Claude Haiku 4.5)

Sources

What Refrase Does

Here is exactly how Refrase optimizes prompts for Claude Haiku 4.5, rule by rule:

XML tag structuring

Refrase wraps your prompt sections in XML tags like <instructions>, <context>, and <output_format>. Claude follows XML-structured prompts more precisely, treating each tag as a distinct instruction block.

Role extraction

Refrase pulls role definitions (like 'You are a data analyst') out of the body text and places them at the top of the prompt in a dedicated <role> block, which Claude uses to calibrate its persona more consistently.

Output format guidance

Refrase adds explicit output format instructions inside an <output_format> tag, ensuring Claude knows exactly what structure to produce (JSON, markdown, bullet list, etc.).

Before / After

See how Refrase transforms a generic prompt for Claude Haiku 4.5.

Original

Extract the key information from this document. Be accurate.

Adapted for Claude Haiku 4.5

<instructions>
<role>You are a precise document extraction specialist.</role>

Extract the key information from this document.

<output_format>Return a JSON object with keys: title, date, author, summary, key_points[].</output_format>
</instructions>

Try It

Your prompt134 chars

Optimized for claude-haiku

Click "Refrase It" or select a model to see the optimized prompt.