Claude Haiku 4.5 vs DeepSeek V3.2 for Constrained Rewriting

Constrained Rewriting (defined in our suite as compression within hard character limits) requires: precise length control, deterministic adherence to output schema, faithfulness to the source, robust paraphrasing that preserves meaning, and reliable handling of long inputs. In our testing there is no third-party external benchmark for this task, so we use our internal scores as the primary signal. Key data points from our tests: constrained_rewriting — DeepSeek V3.2: 4/5; Claude Haiku 4.5: 3/5. Structured_output (JSON/schema compliance) is 5/5 for DeepSeek vs 4/5 for Haiku, explaining why DeepSeek produces outputs that more reliably meet strict character and format limits. Both models score 5/5 on faithfulness and long_context, so neither sacrifices accuracy or long-input retrieval. Tool_calling differs (Haiku 5/5, DeepSeek 3/5), which matters when the rewriting pipeline includes function calls, automated length-testing, or external validators. Context windows: Haiku supports 200,000 tokens vs DeepSeek's 163,840, giving Haiku an edge when compressing very long multimodal documents. All numeric claims cite our internal 1–5 test scores and task ranks.

1. Tight single-sentence compression to a hard character limit (e.g., compress a 400-char passage to 120 characters for a UI badge): DeepSeek V3.2 (constrained_rewriting 4/5, structured_output 5/5) is most likely to hit the limit while preserving meaning. 2) Multi-field JSON output where each field has separate character caps (e.g., title 60 chars, summary 140 chars): DeepSeek's structured_output 5/5 vs Haiku 4/5 favors DeepSeek for schema adherence. 3) Workflows that require calling a length-check tool or chaining functions (e.g., iterative shorten→validate→tune): Claude Haiku 4.5 is preferable because tool_calling is 5/5 for Haiku vs 3/5 for DeepSeek. 4) Compressing content extracted from long documents or images: both models score 5/5 on long_context and faithfulness, but Haiku's larger context window (200,000 tokens) and multimodal input (text+image->text) make it a better fit when the source is very large or contains images. 5) Cost-sensitive bulk rewriting (batch of short snippets): DeepSeek is much cheaper (output cost_per_mtok $0.38 vs Claude Haiku $5.00), which combined with its stronger constrained_rewriting/structured_output scores makes it the pragmatic choice for high-volume tight-compression jobs.

For Constrained Rewriting, choose Claude Haiku 4.5 if you need multimodal input (text+image->text), frequent or complex tool-calling in the rewrite pipeline, or the largest context window (200,000 tokens). Choose DeepSeek V3.2 if your priority is reliably hitting hard character limits and strict schema/format compliance — it wins our constrained_rewriting test by 1 point (4 vs 3) and has stronger structured_output (5/5 vs 4/5) at a much lower output cost ($0.38 per mTok vs $5.00 per mTok).