R1 0528 vs GPT-5.4 for Constrained Rewriting

Constrained Rewriting (defined in our benchmarks as "Compression within hard character limits") demands precise length control, faithfulness to the source, structured-output compliance when a format is required, and predictable completion behavior under tight token budgets. In our testing both R1 0528 and GPT-5.4 scored 4/5 on constrained_rewriting and are tied at rank 6 of 52, so raw task accuracy was comparable. Supporting signals matter when choosing between them: GPT-5.4 scores higher on structured_output (5 vs R1's 4) and safety_calibration (5 vs 4), which helps ensure format-adherent, policy-safe truncations and refusals. R1 0528 shows strong tool_calling (5) and classification (4) support and a 163,840-token context window, but its quirks — notably "empty_on_structured_output" and explicit note that it "Returns empty responses on structured_output, constrained_rewriting, and agentic_planning" and that it "needs_high_max_completion_tokens" — are directly relevant failures for constrained rewriting tasks that rely on short, exact outputs.

GPT-5.4 (winner): Rewriting a legal paragraph to a 280-character SMS while preserving mandatory clauses and returning a JSON flag for omitted clauses. In our testing GPT-5.4's structured_output 5/5 and safety_calibration 5/5 make it reliable for strict format and compliance needs. It also has a 1,050,000-token context window for large source documents. R1 0528 (cost-efficient alternative): Bulk-compressing long product descriptions across multiple languages where you can supply high max-completion tokens and tolerate reasoning-token overhead — R1 0528 is much cheaper (input $0.50/mTok, output $2.15/mTok) and scored 5/5 on tool_calling and 5/5 on persona_consistency/multilingual in our tests. However, in constrained rewriting workflows that expect short, deterministic outputs, R1 0528 may return empty outputs unless you configure very high max_completion tokens (it documents min_max_completion_tokens: 1000) and handle its reasoning-token consumption, making it risky for one-shot short compression tasks.

For Constrained Rewriting, choose R1 0528 if you need a low-cost option for large-scale, multilingual compression and you can set high max_completion_tokens and tolerate reasoning-token output behaviors. Choose GPT-5.4 if you need reliable, format-adherent, and policy-safe compressed outputs out of the box — GPT-5.4 avoids R1 0528's empty-output quirk and has stronger structured-output and safety signals.