Claude Sonnet 4.6 vs GPT-5.4 for Coding

What Coding demands: code correctness, exact schema/format output (JSON or test harnesses), reliable tool calling (tests, linters, repo actions), long-context reasoning across multi-file codebases, faithfulness to source code, and the ability to compress or rewrite code under constraints. Primary benchmark evidence: on SWE-bench Verified (Epoch AI) — the authoritative external test in our payload — GPT-5.4 scores 76.9% vs Claude Sonnet 4.6's 75.2%. That external gap (1.7 points) is small but decisive here. Supporting internal signals (our 1–5 proxies): structured_output (JSON/schema) is GPT-5.4 5 vs Sonnet 4; tool_calling (function selection, argument accuracy, sequencing) is Sonnet 5 vs GPT-5.4 4; constrained_rewriting is GPT-5.4 4 vs Sonnet 3; long_context and faithfulness tie at 5 for both. Also note context windows and costs in the payload: Sonnet 4.6 context_window 1,000,000 and input cost 3 per mTOK; GPT-5.4 context_window 1,050,000 and input cost 2.5 per mTOK (both have output cost 15 per mTOK). These concrete scores explain why GPT-5.4 narrowly wins overall while Sonnet is preferable for some engineering workflows.

1. Single-function correctness and test-passing code generation: GPT-5.4 wins on SWE-bench (76.9% vs 75.2%) and has the higher structured_output score (5 vs 4) — better for exact JSON responses, unit-test-ready snippets, and API payloads. 2) Multi-file refactor with automated tool steps (run tests, apply patch, open PR): Claude Sonnet 4.6 shines in our testing at tool_calling (5 vs GPT-5.4's 4) and agentic planning (both 5), making Sonnet stronger at accurate function selection and argument sequencing for tool-driven workflows. 3) Tight-character/minified solutions or hard compression tasks: GPT-5.4 leads on constrained_rewriting (4 vs 3), so it handles compact rewrites and concise submissions better. 4) Very long-context codebases and large patch generation: both models score 5 on long_context and offer 1M+ token windows (Sonnet 1,000,000; GPT-5.4 1,050,000), so either is suitable for large-repo tasks. 5) Cost-sensitive batch input (many files sent as context): GPT-5.4 has a lower input cost (2.5 vs Sonnet's 3 per mTOK) while output costs match (15 per mTOK), which impacts large-context workflows.

For Coding, choose Claude Sonnet 4.6 if you prioritize tool calling, iterative multi-step dev workflows, and integration with agentic sequences (Sonnet: tool_calling 5 in our tests). Choose GPT-5.4 if you want the slightly higher external correctness on SWE-bench Verified (76.9% vs 75.2%, Epoch AI), stronger structured-output/schema fidelity (5 vs 4), better constrained rewriting (4 vs 3), and marginally lower input cost (2.5 vs 3 per mTOK).