GPT-6 Luna vs GLM 5.3: Benchmark Comparison
GPT-6 Luna has the higher score on 5 of 18 shared benchmarks; GLM 5.3 leads on 13.
The largest observed score gap is 25.25 pts on Terminal-Bench 4.0 , where GLM 5.3 leads.
Reported ±1 standard-error ranges overlap on 9 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 | GPT-6 Luna | GLM 5.3 | Gap | Reported uncertainty |
|---|---|---|---|---|
| Vals Index | 51.22% ±1.07 | 53.51% ±1.30 | 2.30 pts | Reported ±1 SE ranges overlap |
| Harvey's Legal Agent Benchmark | 2.92% ±1.36 | 8.33% ±2.00 | 5.42 pts | Reported ±1 SE ranges do not overlap |
| Legal Research Bench | 30.29% ±3.19 | 49.04% ±3.48 | 18.75 pts | Reported ±1 SE ranges do not overlap |
| EMB | 68.52% ±2.81 | 56.34% ±3.32 | 12.17 pts | Reported ±1 SE ranges do not overlap |
| Finance Agent (v2) | 49.87% ±0.23 | 55.84% ±2.07 | 5.97 pts | Reported ±1 SE ranges do not overlap |
| Tax Agent Bench | 58.86% ±3.21 | 73.09% ±2.97 | 14.23 pts | Reported ±1 SE ranges do not overlap |
| MedCode | 44.69% ±2.30 | 42.86% ±2.11 | 1.82 pts | Reported ±1 SE ranges overlap |
| MedScribe | 83.71% ±1.95 | 88.81% ±2.00 | 5.10 pts | Reported ±1 SE ranges do not overlap |
| ProofBench v1.1 | 64.00% ±4.82 | 49.00% ±5.02 | 15.00 pts | Reported ±1 SE ranges do not overlap |
| MysteryMechanism | 19.37% ±2.66 | 22.97% ±2.83 | 3.60 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench Science | 4.29% ±2.44 | 5.71% ±2.79 | 1.43 pts | Reported ±1 SE ranges overlap |
| Code Migration | 42.55% ±4.42 | 44.22% ±4.29 | 1.67 pts | Reported ±1 SE ranges overlap |
| IOI | 55.56% ±8.87 | 68.44% ±7.48 | 12.89 pts | Reported ±1 SE ranges overlap |
| ProgramBench | 0.50% ±0.50 | 1.50% ±0.86 | 1.00 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench 4.0 | 13.64% ±1.51 | 38.89% ±1.82 | 25.25 pts | Reported ±1 SE ranges do not overlap |
| Vibe Code Bench v1.1 | 81.65% ±3.38 | 78.13% ±3.94 | 3.52 pts | Reported ±1 SE ranges overlap |
| CyberBench v1.1 | 76.25% ±5.29 | 72.08% ±5.41 | 4.17 pts | Reported ±1 SE ranges overlap |
| Public Benefits Bench v1.1 | 57.65% ±1.28 | 68.54% ±1.21 | 10.89 pts | Reported ±1 SE ranges do not overlap |
Performance by category
| Category | GPT-6 Luna average | GLM 5.3 average |
|---|---|---|
| Legal | 16.60% | 28.69% |
| Finance | 59.08% | 61.76% |
| Healthcare | 64.20% | 65.84% |
| Math | 64.00% | 49.00% |
| Science | 11.83% | 14.34% |
| Coding | 38.78% | 46.24% |
| Cyber | 76.25% | 72.08% |
| Social Mobility | 57.65% | 68.54% |
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 | GLM 5.3 cost | GPT-6 Luna latency | GLM 5.3 latency |
|---|---|---|---|---|
| Vals Index | $0.43 | $7.25 | 30m34s | 1h13m |
| Harvey's Legal Agent Benchmark | $0.30 | $4.30 | 17m32s | 42m38s |
| Legal Research Bench | $0.44 | $2.24 | 42m28s | 51m50s |
| EMB | $0.14 | $3.79 | 17m21s | 36m12s |
| Finance Agent (v2) | $0.12 | $1.07 | 14m58s | 15m50s |
| Tax Agent Bench | $0.19 | $1.80 | 33m16s | 50m53s |
| MedCode | N/A | N/A | 108.77s | 2m49s |
| MedScribe | N/A | N/A | 2m50s | 2m02s |
| ProofBench v1.1 | $0.04 | $2.08 | 7m31s | 42m23s |
| MysteryMechanism | $0.03 | $1.53 | 6m01s | 30m42s |
| Terminal-Bench Science | $0.24 | $15.63 | 1h16m | 2h45m |
| Code Migration | $0.60 | $24.91 | 50m23s | 3h47m |
| IOI | $0.20 | $7.67 | 35m47s | 1h46m |
| ProgramBench | $0.18 | $21.96 | 28m55s | 4h16m |
| Terminal-Bench 4.0 | $0.35 | $9.37 | 32m46s | 1h23m |
| Vibe Code Bench v1.1 | $1.35 | $12.45 | 37m00s | 1h04m |
| CyberBench v1.1 | $0.13 | $2.69 | 15m50s | 30m18s |
| Public Benefits Bench v1.1 | $0.22 | $0.95 | 52m59s | 35m51s |
Results available only for GPT-6 Luna
- BioMysteryBench
- SAGE
Results available only for GLM 5.3
- Vals RSI Index
- LegalBench
- TaxEval v2
- GPQA Diamond
- MMLU Pro
- LiveCodeBench
- SkillsBench
- SWE-bench
- Vibe Code Bench 1-100