GPT-6 Luna vs Step 5 Preview: Benchmark Comparison
GPT-6 Luna has the higher score on 7 of 17 shared benchmarks; Step 5 Preview leads on 10.
The largest observed score gap is 22.00 pts on ProofBench v1.1 , where GPT-6 Luna leads.
Reported ±1 standard-error ranges overlap on 9 of 17 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 | GPT-6 Luna | Step 5 Preview | Gap | Reported uncertainty |
|---|---|---|---|---|
| Vals Finance Index | 53.05% ±1.43 | 55.43% ±1.44 | 2.38 pts | Reported ±1 SE ranges overlap |
| Vals Science Index | 28.38% ±1.48 | 29.58% ±1.10 | 1.20 pts | Reported ±1 SE ranges overlap |
| Legal Research Bench | 30.29% ±3.19 | 40.38% ±3.41 | 10.10 pts | Reported ±1 SE ranges do not overlap |
| EMB v1.1 | 69.95% ±2.54 | 69.55% ±2.49 | 0.40 pts | Reported ±1 SE ranges overlap |
| Finance Agent v2.1 | 56.84% ±2.12 | 58.90% ±2.05 | 2.06 pts | Reported ±1 SE ranges overlap |
| Tax Agent Bench | 19.95% ±2.87 | 27.27% ±3.11 | 7.33 pts | Reported ±1 SE ranges do not overlap |
| ProofBench v1.1 | 64.00% ±4.82 | 42.00% ±4.96 | 22.00 pts | Reported ±1 SE ranges do not overlap |
| BioMysteryBench | 61.48% ±2.59 | 67.41% ±0.37 | 5.93 pts | Reported ±1 SE ranges do not overlap |
| MysteryMechanism | 19.37% ±2.66 | 18.47% ±2.61 | 0.90 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench-Science | 4.29% ±2.44 | 2.86% ±2.01 | 1.43 pts | Reported ±1 SE ranges overlap |
| Code Migration v1.1 | 46.67% ±3.61 | 47.00% ±3.43 | 0.33 pts | Reported ±1 SE ranges overlap |
| IOI | 55.56% ±8.87 | 38.06% ±6.69 | 17.50 pts | Reported ±1 SE ranges do not overlap |
| SWE-bench Multimodal | 37.50% ±2.21 | 40.83% ±2.25 | 3.33 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench 4.0 | 13.64% ±1.51 | 31.82% ±2.62 | 18.18 pts | Reported ±1 SE ranges do not overlap |
| Vibe Code Bench v1.1 | 81.65% ±3.38 | 70.40% ±4.27 | 11.25 pts | Reported ±1 SE ranges do not overlap |
| SRE Bench | 2.67% ±1.00 | 0.38% ±0.38 | 2.29 pts | Reported ±1 SE ranges do not overlap |
| Public Benefits Bench v1.1 | 57.65% ±1.28 | 58.12% ±1.28 | 0.47 pts | Reported ±1 SE ranges overlap |
Performance by category
| Category | GPT-6 Luna average | Step 5 Preview average |
|---|---|---|
| Legal | 30.29% | 40.38% |
| Finance | 48.91% | 51.91% |
| Math | 64.00% | 42.00% |
| Science | 28.38% | 29.58% |
| Coding | 47.00% | 45.62% |
| Cyber | 2.67% | 0.38% |
| Social Mobility | 57.65% | 58.12% |
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 | GPT-6 Luna cost | Step 5 Preview cost | GPT-6 Luna latency | Step 5 Preview latency |
|---|---|---|---|---|
| Vals Finance Index | $0.15 | $0.82 | 22m01s | 24m42s |
| Vals Science Index | $0.11 | $1.78 | 30m15s | 1h07m |
| Legal Research Bench | $0.44 | $0.84 | 42m28s | 36m08s |
| EMB v1.1 | $0.14 | $1.33 | 17m21s | 43m35s |
| Finance Agent v2.1 | $0.12 | $0.58 | 15m27s | 9m03s |
| Tax Agent Bench | $0.19 | $0.55 | 33m16s | 21m29s |
| ProofBench v1.1 | $0.04 | $0.71 | 7m31s | 38m34s |
| BioMysteryBench | $0.05 | $0.49 | 8m44s | 21m35s |
| MysteryMechanism | $0.03 | $0.60 | 6m01s | 29m20s |
| Terminal-Bench-Science | $0.24 | $4.25 | 1h16m | 2h30m |
| Code Migration v1.1 | $0.83 | $8.61 | 1h00m | 3h17m |
| IOI | $0.20 | $2.03 | 35m47s | 1h20m |
| SWE-bench Multimodal | $0.03 | $0.61 | 5m16s | 26m06s |
| Terminal-Bench 4.0 | $0.35 | $3.11 | 32m46s | 1h58m |
| Vibe Code Bench v1.1 | $1.35 | $5.86 | 37m00s | 1h31m |
| SRE Bench | $1.55 | $3.07 | 1h21m | 1h34m |
| Public Benefits Bench v1.1 | $0.22 | $0.50 | 52m59s | 27m19s |
Results available only for GPT-6 Luna
- Vals Index
- Vals Legal Index
- HLAB-Vals-1.1
- ClaimBench
- MedCode
- MedScribe
- SAGE
- ProgramBench
- CyberBench v1.1
Results available only for Step 5 Preview
None.