GPT-5.6 Sol leads TryAI Canvas Arena drawing benchmark

// 2h agoBENCHMARK EVAL

GPT-5.6 Sol leads TryAI Canvas Arena drawing benchmark

TryAI benchmarked four frontier AI vision models—GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash—across 28 drawing tasks using an open-source colored pencil toolset on a blank digital canvas. The models executed commands to adjust stroke attributes, lay down colored marks, smudge, erase, and self-review their canvas across target image reproductions (the Mona Lisa and Starry Night) and open-ended text prompts. The evaluation highlighted stark differences in tool-calling behavior, efficiency, and cost: GPT-5.6 Sol produced the best visual details for $7.74 across seven drawings, while Claude Fable 5 cost $160.58 due to deep reasoning overhead and excessive steps. Notably, across all models, structural similarity scores peaked mid-run and declined during later iterations, demonstrating that continuous self-reviewing and over-editing degraded output quality.

// ANALYSIS

Interactive stroke-by-stroke drawing benchmarks expose significant cognitive and execution differences between frontier vision models that static image benchmarks miss.

–**GPT-5.6 Sol Leads in Quality & Cost-Efficiency**: Sol set stroke parameters inline, took fewer average steps (29 per drawing), and delivered superior visual detail for under $8 across seven drawings.
–**Claude Fable 5 Incurs Extreme Cost Overhead**: Fable 5 accumulated an estimated $160.58 bill (~20x higher than peers) due to long reasoning loops and heavy tool usage without matching Sol's visual results.
–**Over-Editing Causes Quality Degradation**: In every target reproduction run, models hit their peak Structural Similarity (SSIM) score mid-task and subsequently eroded fidelity with further edits.
–**SSIM Scores Disconnect from Human Perceptual Artistry**: While Gemini 3.6 Flash recorded the highest raw SSIM scores due to macro color placement, human reviewers consistently preferred Sol's fine line work and shading.

// TAGS

llmvisionbenchmarksmultimodalgpt-5.6claude-fable-5geminigroktryaitryai-canvas-arena

DISCOVERED

2h ago

2026-07-22

PUBLISHED

5h ago

2026-07-21

RELEVANCE

8/ 10

AUTHOR

hershyb_

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

BENCHMARK2h ago

Kimi K3 Mirrors Claude Writing Style

A chart analyzing the stylistic similarity of frontier LLMs reveals that while model outputs usually cluster predictably within their developer families (such as Anthropic's Claudes or OpenAI's GPTs), Kimi K3 is a glaring exception. Instead of exhibiting a distinct writing style, Kimi K3's tone and formatting exhibit strong alignment with Anthropic's Claude models, highlighting notable cross-family stylistic resemblance in frontier AI models.

BENCHMARK2h ago

Kimi K3 aligns with Claude style over model clusters

A chart analyzing the stylistic similarity among frontier large language models highlights a surprising trend: while models typically cluster tightly within their own families (such as GPTs with GPTs or Claudes with Claudes), Moonshot AI's Kimi K3 is a glaring exception. Instead of mirroring typical open-weight model behaviors, Kimi K3 displays writing style and tone characteristics remarkably similar to Anthropic's Claude series.

BENCHMARK2h ago

Kimi K3 stylistically aligns with Anthropic Claude family

A stylistic similarity analysis among frontier large language models highlights an unexpected anomaly in Kimi K3's behavior. While major frontier models typically cluster tightly within their respective model families—such as Claudes clustering with Claudes and GPTs with GPTs—Kimi K3 deviates significantly by exhibiting high stylistic similarity to Anthropic's Claude models.