Claude Haiku 4.5 vs Claude Sonnet 4.6 for Constrained Rewriting

What Constrained Rewriting demands: per our benchmark description it is primarily "compression within hard character limits." That requires: faithful content compression (faithfulness), strict schema/format adherence when truncation rules apply (structured_output), consistent handling of long source material (long_context), and sometimes creative rephrasing to preserve nuance in fewer characters (creative_problem_solving). In our testing both Claude Haiku 4.5 and Claude Sonnet 4.6 scored 3/5 on constrained_rewriting and display the same ranked position ("rank 31 of 53 (22 models share this score)"), so the core capacity for compression is equivalent on the constrained task. Supporting signals: both models score 4 on structured_output and 5 on faithfulness and long_context, indicating they preserve source material and handle long inputs well. Differences that matter operationally: Sonnet 4.6 scores higher on safety_calibration (5 vs 2) and creative_problem_solving (5 vs 4), which helps when compressed outputs must avoid policy issues or require inventive condensation; Haiku 4.5 is described as faster and more cost-efficient, which matters for high-volume batch rewriting.

When to pick each model — grounded in our scores and costs:

• Claude Haiku 4.5 (recommended winner for most constrained-rewrite workloads): batch-compressing product descriptions to a 280-character limit for thousands of SKUs. Both models score 3/5 on the task, but Haiku’s output_cost_per_mtok is 5 versus Sonnet’s 15, so per-token budget is ~3x lower while retaining structured_output 4 and faithfulness 5.
• Claude Sonnet 4.6 (recommended when nuance, safety, or extreme context matter): compressing legal disclaimers or medical summaries where refusal rules and subtle safety tradeoffs matter — Sonnet’s safety_calibration is 5 (vs Haiku 2) and creative_problem_solving is 5 (vs 4), so it’s better at safe, inventive condensations. Sonnet also offers a 1,000,000 token context window and max_output_tokens 128,000 versus Haiku’s 200,000 / 64,000, useful when the rewrite must draw on very long source material.
• Edge case: single, highly creative marketing lines where the brief must preserve tone and inventiveness — Sonnet’s creative_problem_solving advantage helps; for large-scale, cost-sensitive pipelines Haiku is the pragmatic choice. Both models tie on constrained_rewriting score (3/5) and share the same ranked display in our data ("rank 31 of 53 (22 models share this score)").

For Constrained Rewriting, choose Claude Haiku 4.5 if you need equivalent compression quality at much lower cost and faster throughput (output_cost_per_mtok 5 vs 15, described as faster/more efficient). Choose Claude Sonnet 4.6 if you must prioritize safety calibration, richer creative condensation, or ultra-long context (safety_calibration 5 vs 2; creative_problem_solving 5 vs 4; context_window 1,000,000 vs 200,000).