Claude Sonnet 4.6 vs GPT-4.1

Head-to-head results (our 12-test suite plus Epoch AI external measures):

• Wins for Claude Sonnet 4.6: creative_problem_solving 5 vs 3 (Claude tied for 1st among 54, GPT rank 30 of 54), safety_calibration 5 vs 1 (Claude tied for 1st of 55; GPT rank 32 of 55), agentic_planning 5 vs 4 (Claude tied for 1st of 54; GPT rank 16 of 54). These translate to safer refusals, stronger goal decomposition and recovery, and more robust non-obvious idea generation in real tasks.
• Win for GPT-4.1: constrained_rewriting 5 vs 3 (GPT tied for 1st of 53; Claude rank 31 of 53). That maps to better compression within hard character limits and superior performance for microcopy and tight-output transformations.
• Ties (equivalent performance in our tests): structured_output 4/4 (both rank 26 of 54), strategic_analysis 5/5 (both tied for 1st), tool_calling 5/5 (both tied for 1st), faithfulness 5/5 (both tied for 1st), classification 4/4 (both tied for 1st), long_context 5/5 (both tied for 1st), persona_consistency 5/5 (both tied for 1st), multilingual 5/5 (both tied for 1st). Practically, that means both models are excellent at long-context retrieval (30K+), keeping persona, faithful sourcing, function selection, and multilingual output.
• External benchmarks (Epoch AI): on SWE-bench Verified (coding) Claude scores 75.2% vs GPT-4.1's 48.5% (Claude rank 4 of 12; GPT rank 11 of 12), indicating a clear edge for Claude on real GitHub issue resolution. On AIME 2025 (math olympiad) Claude scores 85.8% vs GPT-4.1's 38.3% (Epoch AI), an advantage for high‑competency contest math. GPT-4.1 posts 83% on MATH Level 5 (Epoch AI), a strong result where Claude has no reported score in the payload. Overall, Claude wins the plurality of benchmarks in our suite (3 wins vs GPT's 1) and shows much higher external coding and AIME scores; GPT-4.1 keeps parity on many core capabilities and wins the constrained-rewriting specialty.

Payload prices: Claude Sonnet 4.6 = $3 input / $15 output per mTok; GPT-4.1 = $2 input / $8 output per mTok. Interpreting mTok as per million tokens, raw output-only cost is $15 vs $8 for 1M tokens, $150 vs $80 for 10M, and $1,500 vs $800 for 100M. For a 50/50 input/output token split the totals are Claude $9 vs GPT $5 for 1M; $90 vs $50 for 10M; $900 vs $500 for 100M. The upshot: Claude is ~1.875x more expensive overall (priceRatio 1.875). Teams with heavy monthly throughput (10M+ tokens) or tight budgets should favor GPT-4.1 to save several hundred to thousands of dollars monthly; teams that need better safety calibration, agentic planning, or higher external coding scores may justify Claude's premium.

Choose Claude Sonnet 4.6 if you need: safety-first production agents, high-stakes planning or creative problem-solving, and stronger coding performance (SWE-bench Verified 75.2%); the model scores 5/5 on safety_calibration and agentic_planning but costs more ($3/$15 per mTok). Choose GPT-4.1 if you need: the best value for high-volume throughput, superior constrained rewriting (5/5), or a lower-cost generalist that matches Claude on faithfulness, long-context, tool calling and multilingual tasks; GPT pricing is $2/$8 per mTok and it posts 83% on MATH Level 5 (Epoch AI).