Physics Reasoning: implicit optics in AI image models

In PromptFrenzy's 42-run implicit-optics benchmark, Nano Banana 2 passed five of six runs; no other model passed more than one.

Finding

Google Nano Banana 2 was the only tested model to apply the implied optics reliably, passing five of six runs (83.3%); every other model passed at most one of six.

Methodology

Two implicit-optics tests used neutral prompts that never named refraction, reflection or the expected effect. Seven models generated three runs per test with no retries. Each run was hand-graded pass/fail against a disclosed physical criterion; all 42 source images and exact prompts are published. Sample size: 42 runs. Published: 2026-07-13.

Results

1. Google Nano Banana 2 — 83.3/100
2. Meta Muse Image — 16.7/100
3. Nano Banana 2 Lite — 16.7/100
4. FLUX Kontext Pro — 0.0/100
5. OpenAI gpt-image-2 — 0.0/100
6. Qwen Image 2.0 Pro — 0.0/100
7. Seedream 5 Lite — 0.0/100

Download and verify

• Complete JSON dataset
• Run-level CSV
• Every source image and exact prompt

Source: PromptFrenzy Physics Reasoning benchmark, published 13 July 2026. CC BY 4.0; link to the canonical research page when citing.

Run the refraction test yourself → open the benchmark.

Originally published at PromptFrenzy.