GACC-Bench (10% budget) - Macro accuracy: leaderboard
GACC-Bench-full (500 goals, 2,571 questions over 281 contexts from seven long-document and agent-trace collections). The listed model is the goal-conditioned compressor; questions are withheld from it. A fixed DeepSeek-V4-Pro reader answers each question from the compressed context and a fixed DeepSeek-V4-Pro judge scores it against the reference (98.25% human agreement on 114 sampled responses). Retention budget 10% of source-context tokens. Macro accuracy is the percentage of goals for which every associated question is answered correctly, computed over all 500 goals. Model labels as printed; reasoning or effort settings are not disclosed. Full-context reference: 99.38 Micro / 96.80 Macro; no-context: 6.46 / 0.20. Metric: Goal accuracy, all questions correct (%). Higher is better. Source: https://arxiv.org/html/2610.21893v1
Metric: Goal accuracy, all questions correct (%). Source: arxiv.org. Saturation forecast: Around 2031. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 14.8 | #6 |
| 2 | GPT-6.1 Sol | 14.8 | #7 |
| 3 | GLM-5.3 | 10.8 | #40 |
| 4 | GLM-5.3 Flash | 10 | #63 |
| 5 | Claude Opus 5.5 | 9.2 | #4 |
| 6 | Kimi K3 | 6.6 | #29 |
| 7 | Qwen 3.8 Max | 6.4 | #32 |
| 8 | DeepSeek V4 Pro | 6.2 | #102 |
| 9 | GPT-5.6 Sol | 5.2 | #20 |
| 10 | Gemini 3.6 Flash | 3.6 | #51 |
No result here: #8 Gemini 4 Argon, #9 Claude Fable 5.1, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=gacc-bench-10-budget-macro-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-18.