benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $1.00 |
| Standard output | $5.00 |
| Cached input | $0.10 |
| Context window | 200,000 tokens |
| Max output | 64,000 tokens |
How to measure latency without a made-up speed figure
Anthropic's public rate card does not provide a universal tokens-per-second figure for this comparison, and benchr has not run a controlled cross-region latency study. We removed the previous 145 tok/s planning value instead of preserving an undocumented editorial estimate.
Benchmark both models in the target region with the same prompts, output lengths, settings, and concurrency. Record time to first token plus p50, p95, and p99 completion time; those observations are usable for capacity planning, unlike a single undocumented number.
Where the 73.3% SWE-bench score creates limits
Haiku 4.5's provider-published 73.3% SWE-bench score is six points below Sonnet 4.6's 79.6%. That benchmark difference does not translate into “six fewer issues per 100” in production, establish where a gap will be visible, or guarantee either model's success rate. Both models support 64K max output, while Haiku lists 200K context and Sonnet lists 1M. Use a held-out evaluation of your own repositories and failure criteria.
Cost scenarios
At 50M input + 10M output per month — a moderate production volume: Haiku 4.5 costs $50 + $50 = $100/month. Claude Sonnet 4.6 at the same volume: $150 + $150 = $300/month — 3× the cost. With 90% cache hit on Haiku: approximately $9.50 + $50 = $59.50/month. For a product routing 10,000 user queries per day through a classification step, Haiku at $100/month versus Sonnet at $300/month is a $200/month cost difference that scales directly with volume.
Use-case fit
Candidate for: High-volume classification and intent detection; customer service first-response routing; extraction from structured documents; and pipelines where the published rate card is attractive. Validate throughput and task quality before using it for latency-sensitive chat.
Evaluate another tier too if: Your tasks require complex multi-step code generation, architecture reasoning, more than 200K context, or an independently validated latency target. The SWE-bench gap alone does not predict your production pass rate.
Decision checklist
Before choosing Haiku over Sonnet: run a sample of your actual tasks on both models and check for quality degradation in your specific failure modes. If your task set is pure classification or extraction with structured outputs, Haiku is likely sufficient. If it involves code generation or open-ended reasoning, measure pass rates on both before committing.
Check your required output length against Haiku's published 64K max output and your input size against its 200K context window. Sonnet 4.6 has the same 64K max output but expands context to 1M.