Claude Haiku 4.5 vs Claude Opus 4.6 for Constrained Rewriting

What Constrained Rewriting demands: precise length control and strict adherence to a character/byte budget while preserving meaning and required content. Key capabilities: structured_output (schema/format adherence), faithfulness (preserve source facts and intent), long_context (retain full source when long), persona_consistency (if voice must be preserved), and deterministic tool-like behavior for exact truncation or compression. In our testing both Claude Haiku 4.5 and Claude Opus 4.6 score 3/5 on the constrained_rewriting test (rank 31/52 for each). Supporting proxy metrics show parity on the most relevant dimensions: structured_output = 4 (both), faithfulness = 5 (both), long_context = 5 (both), and tool_calling = 5 (both). Differences that explain real-world tradeoffs: Opus 4.6 has much higher safety_calibration (5 vs 2) and higher creative_problem_solving (5 vs 4), while Haiku 4.5 is faster/more efficient and far cheaper per output token (input/output cost per m-token: Haiku 1/5 vs Opus 5/25). Because the primary task score is tied, these supporting dimensions drive the recommendation.

1. High-volume UI copy compression (same rules for every item): Use Claude Haiku 4.5. Both models score 3/5 on constrained_rewriting and share structured_output 4 and faithfulness 5, but Haiku costs $5 per output m-token vs Opus $25 — materially cheaper for batch jobs. 2) Safety-critical legal redaction into a strict character limit: Use Claude Opus 4.6. Both score 3/5, but Opus’s safety_calibration is 5 vs Haiku 2 in our testing, reducing risk when content must be refused or strictly sanitized. 3) Creative newsletter compression where preserving tone while shortening massively is needed: Prefer Opus 4.6 — creative_problem_solving 5 vs Haiku 4 gives Opus an advantage for inventive rephrasing within tight limits. 4) Low-latency in-product shortening (on-device or cost-sensitive API): Prefer Haiku 4.5 for its efficiency and lower output cost while retaining equal faithfulness and structured-output compliance in our tests.

For Constrained Rewriting, choose Claude Haiku 4.5 if you need equivalent compression quality at much lower cost and higher throughput (both score 3/5; Haiku output cost $5 vs Opus $25 per m-token). Choose Claude Opus 4.6 if safety-critical filtering or stronger creative rephrasing matters more than cost (Opus has safety_calibration 5 and creative_problem_solving 5 in our testing).