Claude Sonnet 4.6 vs R1 0528 for Constrained Rewriting

Constrained Rewriting (compression within hard character limits) demands tight length control, high faithfulness to the source, reliable adherence to output constraints (including schema/format when required), and efficient token usage for short outputs. There is no external benchmark for this task in the payload, so our internal constrained_rewriting scores are the primary signal: R1 0528 scores 4/5 (taskRank 6/52) vs Claude Sonnet 4.6 at 3/5 (taskRank 31/52). Supporting signals: both models score 5/5 on faithfulness and 4/5 on structured_output (a tie), meaning both can preserve meaning and follow formats in many cases. Differences that matter operationally: R1's quirks note it “returns empty responses on structured_output, constrained_rewriting, and agentic_planning — reasoning tokens consume output budget on short tasks” and requires high max_completion_tokens (min_max_completion_tokens: 1000). Claude Sonnet 4.6 supports structured_outputs and verbosity in its parameter set and does not list the same empty-response quirk, but it costs substantially more (output_cost_per_mtok $15 vs R1’s $2.15), and its constrained_rewriting score is lower. Use the internal 1–5 task scores and these implementation details to weigh raw capability vs reliability and cost.

1. High-volume push-notification compression to 140 chars: R1 0528 is preferable — it scores 4/5 on constrained_rewriting and has much lower output cost ($2.15 per mTok vs $15), reducing operating expense at scale. Ensure you set max_completion_tokens above R1's min (>=1000) and test for empty outputs. 2) One-off legal boilerplate compressed to a strict template (must produce JSON schema): Claude Sonnet 4.6 is a safer pick if you require guaranteed, non-empty schema-compliant output without tuning, because Sonnet lists structured_outputs in its supported params and does not carry R1’s empty-response quirk — but expect a higher cost (Sonnet output $15 per mTok). 3) Mixed-media source (image + text) compressed into a short caption: Claude Sonnet 4.6 supports text+image->text modality (R1 is text->text), so Sonnet may be necessary if the input includes images; however, for pure text compression, R1 wins on our constrained_rewriting proxy. 4) Tight short-task pipelines where reasoning tokens matter: R1 can consume reasoning tokens and return empties on short tasks; in those automated pipelines you must allocate large max_completion_tokens or avoid structured_output flags, otherwise Sonnet's more predictable behavior is preferable despite its lower constrained_rewriting score.

For Constrained Rewriting, choose R1 0528 if you need higher measured compression performance and much lower output costs (R1 scores 4 vs Sonnet 4.6's 3 on our test and costs $2.15 vs $15 per mTok) and you can configure high max_completion_tokens to avoid empty responses. Choose Claude Sonnet 4.6 if you require guaranteed non-empty, schema-compliant outputs, image->text input support, or you prefer a model without R1’s empty-response quirk and are willing to pay ~7× more per output token.