Benchmark Almanac

LongBanana

OpenClaw + RPR + Gemini 3.1 Flash Image 78.4 tops its leaderboard Gemini 3.1 Flash Image 74.6 leads as a model on its own
Not in model cards

Models synthesize images from many reference images and textual constraints. Atomic rubric checks score content fidelity, relationships, and global constraints across 90 tasks.

Tasks average about 32 references and 20 constraints. The 4,670 checks award 0, 1, or 2 points before normalization per case. The current repository uses an unweighted mean of case scores. Historical weighted paper results require their original protocol.

Released
November 2026
Kind
Image generation and editing
Level
Everyday
Scored
Graded by a model
Answers
Open-ended
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images → Images
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
Last model added
February 2026
Scores refreshed
November 20th, 2026

In the authors’ words

We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.

Abstract of LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis (arXiv:2611.06856, CC0)
9
Entries on its own leaderboard, 1 model we track

OpenClaw + RPR + Gemini 3.1 Flash Image tops the leaderboard with 78.4 Gemini 3.1 Flash Image leads with 74.6

Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran on its own with several settings, we keep its best. Agents other companies built around a model get a pale shade of that model's lab color. Gray marks blends and entries not in our model list, such as fine-tunes and older models. None of these count toward leaders. The best model on its own is Gemini 3.1 Flash Image, at 74.6. Agents built around a model, such as OpenClaw + RPR + Gemini 3.1 Flash Image, are their own entries. We don't count them as the model's score. This view shows only models we track, each at its best.

  1. 1 OpenClaw + RPR + Gemini 3.1 Flash Image 78.4
  2. 2 OpenCode + RPR + Gemini 3.1 Flash Image 77.8
  3. 3 OpenCode + Gemini 3.1 Flash Image 77.3
  4. 4 OpenClaw + Gemini 3.1 Flash Image 76.1
  5. 5 Hermes-Agent + RPR + Gemini 3.1 Flash Image 75.6
  6. 6 Gemini 3.1 Flash Image PD-Direct 74.6
  7. 7 Hermes-Agent + Gemini 3.1 Flash Image 73.9
  8. 8 OpenClaw + RPR + Seedream 4.0 63.6
  9. 9 OpenClaw + RPR + GPT-Image-1 61.8

Source: Official LongBanana leaderboard

Source: Official LongBanana leaderboard

#EntryScoreSettingDateSource
1 OpenClaw + RPR + Gemini 3.1 Flash Image 78.4 Feb 2026 Official leaderboard
2 OpenCode + RPR + Gemini 3.1 Flash Image 77.8 Feb 2026 Official leaderboard
3 OpenCode + Gemini 3.1 Flash Image 77.3 Feb 2026 Official leaderboard
4 OpenClaw + Gemini 3.1 Flash Image 76.1 Feb 2026 Official leaderboard
5 Hermes-Agent + RPR + Gemini 3.1 Flash Image 75.6 Feb 2026 Official leaderboard
6 Gemini 3.1 Flash Image 74.6 PD-Direct Feb 2026 Official leaderboard
7 Hermes-Agent + Gemini 3.1 Flash Image 73.9 Feb 2026 Official leaderboard
8 OpenClaw + RPR + Seedream 4.0 63.6 Official leaderboard
9 OpenClaw + RPR + GPT-Image-1 61.8 Official leaderboard
#ModelScoreSettingDateSource
1 Gemini 3.1 Flash Image 74.6 PD-Direct Feb 2026 Official leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard
68.0 70.0 72.0 74.0 76.0 2026 68.0 70.0 72.0 74.0 76.0 2026

Source: Official LongBanana leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Vision

All 636 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. Qwen 3.6 Plus 86%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy. Seed2.1 Pro 90.7%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%

Data audit

8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 9 months ago, in February 2026, for Gemini 3.1 Flash Image. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the official leaderboard (9 entries).
  • Pass Models scored 1 model we track, plus 8 other entries on its leaderboard.
  • Warning Newest score The newest came 9 months ago, in February 2026, for Gemini 3.1 Flash Image.
  • Note Frontier models None of its models are on the capabilities index.
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on klingairesearch.github.io: 10 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check