Ember-1 vs GPT-5.6 Luna: Benchmark Comparison
Ember-1 has the higher score on 10 of 12 shared benchmarks; GPT-5.6 Luna leads on 2.
The largest observed score gap is 15.45 pts on IOI , where GPT-5.6 Luna leads.
Reported ±1 standard-error ranges overlap on 4 of 12 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 | Ember-1 | GPT-5.6 Luna | Gap | Reported uncertainty |
|---|---|---|---|---|
| Vals Science Index | 31.76% ±1.66 | 25.30% ±0.82 | 6.46 pts | Reported ±1 SE ranges do not overlap |
| Legal Research Bench | 44.71% ±3.46 | 36.54% ±3.35 | 8.17 pts | Reported ±1 SE ranges do not overlap |
| Tax Agent Bench | 31.08% ±3.34 | 24.82% ±3.11 | 6.26 pts | Reported ±1 SE ranges overlap |
| ProofBench v1.1 | 61.00% ±4.90 | 60.00% ±4.92 | 1.00 pts | Reported ±1 SE ranges overlap |
| BioMysteryBench | 67.04% ±3.03 | 61.48% ±0.74 | 5.56 pts | Reported ±1 SE ranges do not overlap |
| MysteryMechanism | 22.52% ±2.81 | 14.41% ±2.36 | 8.11 pts | Reported ±1 SE ranges do not overlap |
| Terminal-Bench-Science | 5.71% ±2.79 | 0.00% ±0.00 | 5.71 pts | Reported ±1 SE ranges do not overlap |
| IOI | 46.33% ±9.16 | 61.78% ±11.61 | 15.45 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench 4.0 | 20.71% ±3.64 | 11.62% ±1.01 | 9.09 pts | Reported ±1 SE ranges do not overlap |
| Vibe Code Bench v1.1 | 83.87% ±3.15 | 77.06% ±3.07 | 6.81 pts | Reported ±1 SE ranges do not overlap |
| CyberBench v1.1 | 69.52% ±5.56 | 73.63% ±5.56 | 4.11 pts | Reported ±1 SE ranges overlap |
| Public Benefits Bench v1.1 | 66.17% ±1.23 | 61.16% ±1.27 | 5.01 pts | Reported ±1 SE ranges do not overlap |
Performance by category
| Category | Ember-1 average | GPT-5.6 Luna average |
|---|---|---|
| Legal | 44.71% | 36.54% |
| Finance | 31.08% | 24.82% |
| Math | 61.00% | 60.00% |
| Science | 31.76% | 25.30% |
| Coding | 50.30% | 50.15% |
| Cyber | 69.52% | 73.63% |
| Social Mobility | 66.17% | 61.16% |
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 | Ember-1 cost | GPT-5.6 Luna cost | Ember-1 latency | GPT-5.6 Luna latency |
|---|---|---|---|---|
| Vals Science Index | $10.99 | $0.26 | 1h18m | 50m00s |
| Legal Research Bench | $3.67 | $0.85 | 38m34s | 39m34s |
| Tax Agent Bench | $2.97 | $0.47 | 30m46s | 42m46s |
| ProofBench v1.1 | $1.74 | $0.12 | 21m09s | 8m34s |
| BioMysteryBench | $0.91 | $0.09 | 9m26s | 18m31s |
| MysteryMechanism | $1.29 | $0.11 | 20m15s | 8m13s |
| Terminal-Bench-Science | $30.76 | $0.58 | 3h23m | 2h03m |
| IOI | $24.46 | $0.58 | 1h47m | 1h10m |
| Terminal-Bench 4.0 | $19.48 | $0.73 | 1h55m | 35m55s |
| Vibe Code Bench v1.1 | $8.56 | $0.73 | 39m18s | 25m52s |
| CyberBench v1.1 | $2.43 | $0.40 | 22m02s | 16m15s |
| Public Benefits Bench v1.1 | $1.02 | $0.27 | 13m21s | 38m40s |
Results available only for Ember-1
None.
Results available only for GPT-5.6 Luna
- LegalBench
- MortgageTax
- TaxEval v2
- MedCode
- MedScribe
- GPQA Diamond
- MMLU Pro
- MMMU Pro
- SAGE
- ProgramBench
- SkillsBench
- SWE-bench
- Vibe Code Bench 1-100
- SRE Bench