LongBanana - P70: leaderboard
Reference-grounded image synthesis with explicit image generator and controller identities. LongBanana v1 uses 90 historical cases, 4,670 atomic checklist items, 900 generated outputs and 46,700 reviewed human labels. Every agentic condition uses Claude Sonnet 4.6 as its controller. Table 2 fixes Nano Banana 2 and varies the workflow; Table 3 fixes OpenClaw plus RPR and varies the image generator. FP-Direct, PD-Direct, each harness and each RPR workflow retain distinct identities. Category and Overall columns use unweighted checklist averages on 0–100; P@70 uses the native difficulty-weighted task pass rate. The repeated Nano Banana 2/OpenClaw-plus-RPR condition is emitted once. These results are not relabeled for open_v1p5, current repository API defaults or Appendix Table 11’s differently weighted ablations. Question access: The public open_v1 release contains 90 cases, 2,880 references and 4,670 checks, matching the historical cohort. open_v1p5 is a different release and is not substituted.
Metric: Difficulty-weighted task pass (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=OpenCode; controller=Claude Sonnet 4.6; historical v1) | 47.8 | |
| 2 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=OpenClaw + RPR; controller=Claude Sonnet 4.6; historical v1) | 47.8 | |
| 3 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=Hermes-Agent; controller=Claude Sonnet 4.6; historical v1) | 43.3 | |
| 4 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=OpenCode + RPR; controller=Claude Sonnet 4.6; historical v1) | 43.3 | |
| 5 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=OpenClaw; controller=Claude Sonnet 4.6; historical v1) | 40 | |
| 6 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=PD-Direct; controller=none; historical v1) | 35.6 | |
| 7 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=Hermes-Agent + RPR; controller=Claude Sonnet 4.6; historical v1) | 35.6 | |
| 8 | Nano Banana 2 (Benchmark Sweep Run: source=LongBanana; workflow=FP-Direct; controller=none; historical v1) | 24.4 | |
| 9 | Seedream 4 (Benchmark Sweep Run: source=LongBanana; workflow=OpenClaw + RPR; controller=Claude Sonnet 4.6; historical v1) | 18.9 | |
| 10 | GPT Image 1 (Benchmark Sweep Run: source=LongBanana; workflow=OpenClaw + RPR; controller=Claude Sonnet 4.6; historical v1) | 17.8 |
Interactive version: theaggregate.ai/benchmark?slug=longbanana-p70 · How It Works · Data refreshed daily, snapshot 2026-11-06.