Model comparison
Claude Sonnet 5 vs GPT-6 Luna: Benchmark Comparison
Claude Sonnet 5 has the higher score on 10 of 18 shared benchmarks; GPT-6 Luna leads on 8.
The largest observed score gap is 14.34 pts on CyberBench v1.1 , where GPT-6 Luna leads.
Reported ±1 standard-error ranges overlap on 11 of 18 shared benchmarks with comparable uncertainty data. This is not a pairwise statistical significance test.
Shared benchmark results
Scores and reported standard errors come from the latest visible Vals benchmark version available for both models.
| Benchmark | Claude Sonnet 5 | GPT-6 Luna | Gap | Reported uncertainty |
|---|---|---|---|---|
| Vals Index | 51.77% ±1.09 | 51.22% ±1.07 | 0.56 pts | Reported ±1 SE ranges overlap |
| Harvey's Legal Agent Benchmark | 5.00% ±1.65 | 2.92% ±1.17 | 2.08 pts | Reported ±1 SE ranges overlap |
| Legal Research Bench | 41.83% ±3.43 | 30.29% ±3.19 | 11.54 pts | Reported ±1 SE ranges do not overlap |
| EMB | 66.32% ±3.01 | 68.52% ±2.81 | 2.20 pts | Reported ±1 SE ranges overlap |
| Finance Agent (v2) | 53.91% ±0.52 | 49.87% ±0.23 | 4.04 pts | Reported ±1 SE ranges do not overlap |
| Tax Agent Bench | 62.27% ±3.19 | 58.86% ±3.21 | 3.41 pts | Reported ±1 SE ranges overlap |
| MedCode | 47.54% ±2.27 | 44.69% ±2.30 | 2.85 pts | Reported ±1 SE ranges overlap |
| MedScribe | 76.05% ±3.05 | 83.71% ±1.95 | 7.66 pts | Reported ±1 SE ranges do not overlap |
| ProofBench v1.1 | 77.00% ±4.23 | 64.00% ±4.82 | 13.00 pts | Reported ±1 SE ranges do not overlap |
| Terminal-Bench Science | 2.86% ±2.01 | 4.29% ±2.44 | 1.43 pts | Reported ±1 SE ranges overlap |
| SAGE | 48.92% ±3.40 | 48.09% ±3.38 | 0.83 pts | Reported ±1 SE ranges overlap |
| Code Migration | 44.39% ±4.25 | 42.55% ±4.42 | 1.84 pts | Reported ±1 SE ranges overlap |
| IOI | 45.00% ±2.75 | 55.56% ±8.87 | 10.56 pts | Reported ±1 SE ranges overlap |
| ProgramBench | 0.00% ±0.00 | 0.50% ±0.50 | 0.50 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench 4.0 | 9.60% ±1.01 | 13.64% ±1.51 | 4.04 pts | Reported ±1 SE ranges do not overlap |
| Vibe Code Bench v1.1 | 81.33% ±3.05 | 81.65% ±3.38 | 0.32 pts | Reported ±1 SE ranges overlap |
| CyberBench v1.1 | 61.91% ±5.74 | 76.25% ±5.29 | 14.34 pts | Reported ±1 SE ranges do not overlap |
| Public Benefits Bench v1.1 | 66.03% ±1.23 | 57.65% ±1.28 | 8.39 pts | Reported ±1 SE ranges do not overlap |
Performance by category
| Category | Claude Sonnet 5 average | GPT-6 Luna average |
|---|---|---|
| Legal | 23.41% | 16.60% |
| Finance | 60.83% | 59.08% |
| Healthcare | 61.80% | 64.20% |
| Math | 77.00% | 64.00% |
| Science | 2.86% | 4.29% |
| Education | 48.92% | 48.09% |
| Coding | 36.06% | 38.78% |
| Beta | 61.91% | 76.25% |
| Social Mobility | 66.03% | 57.65% |
Cost and latency
Cost per test appears only where the benchmark reports it for both models. Latency is the measured completion time for that benchmark.
| Benchmark | Claude Sonnet 5 cost | GPT-6 Luna cost | Claude Sonnet 5 latency | GPT-6 Luna latency |
|---|---|---|---|---|
| Vals Index | $13.72 | $0.43 | 54m16s | 30m34s |
| Harvey's Legal Agent Benchmark | $8.95 | $0.30 | 38m52s | 17m32s |
| Legal Research Bench | $2.72 | $0.44 | 25m46s | 42m28s |
| EMB | $10.29 | $0.14 | 49m46s | 17m21s |
| Finance Agent (v2) | $0.75 | $0.12 | 13m12s | 14m58s |
| Tax Agent Bench | $1.79 | $0.19 | 19m19s | 33m16s |
| MedCode | N/A | N/A | 2m15s | 108.77s |
| MedScribe | N/A | N/A | 4m12s | 2m50s |
| ProofBench v1.1 | $1.37 | $0.04 | 14m58s | 7m31s |
| Terminal-Bench Science | $18.70 | $0.24 | 2h55m | 1h16m |
| SAGE | N/A | N/A | 7m14s | 91.07s |
| Code Migration | $35.31 | $0.60 | 1h57m | 50m23s |
| IOI | $12.89 | $0.20 | 59m44s | 35m47s |
| ProgramBench | $24.36 | $0.18 | 1h31m | 28m55s |
| Terminal-Bench 4.0 | $26.33 | $0.35 | 1h45m | 32m46s |
| Vibe Code Bench v1.1 | $25.39 | $1.35 | 1h07m | 37m00s |
| CyberBench v1.1 | $1.89 | $0.13 | 16m57s | 15m50s |
| Public Benefits Bench v1.1 | $1.29 | $0.22 | 21m09s | 52m59s |
Results available only for Claude Sonnet 5
- LegalBench
- MortgageTax
- TaxEval v2
- GPQA Diamond
- MMLU Pro
- MMMU Pro
- LiveCodeBench
- SkillsBench
- SWE-bench
- Vibe Code Bench 1-100
Results available only for GPT-6 Luna
- BioMysteryBench
- MysteryMechanism