Claude Haiku 4.5 vs Devstral Medium for Constrained Rewriting

What Constrained Rewriting demands: tight compression with zero tolerance for dropped meaning, strict adherence to character limits and output format, and the ability to handle long or multi-part sources. Key capabilities: faithfulness (preserve original meaning), structured_output (adhere to schemas and exact formats), long_context (process long inputs without losing earlier context), and tool_calling (accurate function selection/arguments for length checks and iterative trimming). In our testing the primary signal for this task is the constrained_rewriting score — both models score 3/5, tying them on the core test. Use supporting metrics to break the tie: Claude Haiku 4.5 scores higher on long_context (5 vs 4), faithfulness (5 vs 4), and tool_calling (5 vs 3); structured_output is equal at 4 for both. Those supporting differences explain why Haiku produces more reliable, faithful compressions on long or format-sensitive inputs even though the headline task score is identical.

High-stakes long-document compression: Claude Haiku 4.5 is better for compressing long legal or academic paragraphs into strict character-limited summaries because long_context 5 (Haiku) vs 4 (Devstral) and faithfulness 5 vs 4 reduce drift and preserve nuance. Function-driven pipelines: Haiku's tool_calling 5 vs 3 makes it preferable when your agent calls a char-count/check function iteratively to meet hard limits. High-volume short rewrites with tight budgets: Devstral Medium is attractive when constrained_rewriting quality requirements are moderate — it ties 3/5 with Haiku but costs less (input_cost_per_mtok 0.4 vs 1; output_cost_per_mtok 2 vs 5), so you can run many short rewrites cheaply. Format-sensitive outputs (JSON/strict templates): both models score 4/5 on structured_output in our tests, so either works for schema-adherent compressed outputs; prefer Haiku if the source is long or meaning preservation is critical.

For Constrained Rewriting, choose Claude Haiku 4.5 if you need the most reliable, faithful compression for long or format-sensitive sources (long_context 5 vs 4; faithfulness 5 vs 4; tool_calling 5 vs 3) and you can accept higher costs (input 1 / output 5 per mTok). Choose Devstral Medium if budget per call is the priority and your rewrites are short or lower-risk — it ties on the core task (3/5) but is cheaper (input 0.4 / output 2 per mTok).