§ Ranked #05 by discriminative power
GBA Eval.
An environment for long-horizon SWE (build a GBA emulator). Across 10 models with public results it spreads the best and worst 53% — it still sorts frontier models, so training on it can still move yours.
§ Public model scores
Who wins GBA Eval.
Best public result per model entry, normalized 0..1. The spread between the top and bottom rows is what makes this environment worth — or not worth — a training run.
| # | Model | overall |
|---|---|---|
| 01 | GPT-5.5 | 53% |
| 02 | Sonnet-4.6 | 49% |
| 03 | Opus-4.6 | 44% |
| 04 | Opus-4.7 | 44% |
| 05 | GPT-5.4 | 32% |
| 06 | Gemini-3.5-Flash | 7% |
| 07 | Kimi-K2.6 | 1% |
| 08 | Gemini-3.1-Pro | 1% |
| 09 | GLM-5.1 | 0% |
| 10 | MiniMax-M2.7 | 0% |
§ Nearby in the ranking
| # | Environment | Spread | Discriminative |
|---|---|---|---|
| 03 | DeepSWElong-horizon agentic coding | 65% | 0.65 |
| 04 | FrontierSWEfrontier software engineering | 59% | 0.59 |
| 05 | GBA Evallong-horizon SWE (build a GBA emulator) | 53% | 0.53 |
| 06 | SkillsBenchagent skills (multi-domain) | 43% | 0.43 |
| 07 | Diplomacy Arenastrategy / negotiation (games) | 29% | 0.29 |
§ Work with us
Need one that still separates models?
When the public environment for your capability saturates, you can’t tell your models apart and you can’t train past it. We build private, contamination-resistant, verifiable-reward environments and evals on a hold-out set — designed to discriminate where the public ones no longer do.