Claude 3.5 Haiku
Anthropic
Anthropic's fastest model that matches Claude 3 Opus performance at similar speed to previous-gen Haiku. Released October 2024 with July 2024 training cutoff. Achieves 40.6% on SWE-bench Verified, outperforming agents using Claude 3.5 Sonnet and GPT-4o. Features 200K context window and 8K max output. Surpasses Claude 3 Opus on intelligence benchmarks while maintaining Haiku-class speed. Pricing: $0.80 per 1M input, $4 per 1M output (revised Dec 2024). Supports prompt caching (90% savings) and Message Batches API (50% savings).
Strengths
- Matches Claude 3 Opus intelligence at Haiku-class speed - best of both worlds
- 40.6% on SWE-bench Verified outperforms Claude 3.5 Sonnet and GPT-4o agents
- 200K context window for large document processing
- 8K max output tokens (2x larger than Claude 3 Haiku's 4K)
- July 2024 training cutoff more recent than many competitors
- Prompt caching enables up to 90% cost savings for repetitive tasks
- Message Batches API provides 50% additional savings
Caveats
- No vision capabilities (text-only) unlike Claude 3 Haiku
- More expensive than Claude 3 Haiku ($0.80 vs $0.25 input)
- Pricing increased from initial $1 to $0.80 after revision
- Lower ceiling than Claude 3.5 Sonnet or Claude Opus 4.5 for complex tasks
- 8K max output still lower than GPT-5 series (64K-128K)
- Vendor
- Anthropic
- Context Window
- 200,000 tokens
- Max Output
- 8,192 tokens
- License
- Proprietary
- Release Date
- 2024-10-22
- Training Cutoff
- 2024-07-01
- Modalities
- text
| Benchmark | Score | Max |
|---|---|---|
| SWE-bench Verified | 40.6 | - |
| livebench_coding | 51.4 | 100.0 |
| livebench_language | 39.1 | 100.0 |
| livebench_instruction_following | 68.8 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
$0.80 per 1M input tokens, $4.00 per 1M output tokens. Up to 90% savings with prompt caching