inferwire
/
AI·4 min read

Mona Lisa Pencil Test: Sol Judged Best, Fable 5 Cost the Most

TryAI had four AI models draw with colored-pencil tools on a blank canvas. Its team rated GPT-5.6 Sol best, Gemini 3.6 Flash scored highest on similarity, and Claude Fable 5 cost the most.

TL;DR

  • TryAI gave four models colored-pencil drawing tools; its team rated GPT-5.6 Sol best, while Gemini 3.6 Flash scored highest on image similarity [^1].
  • Claude Fable 5 spent an estimated $160.58 on seven drawings, against $7.74 to $12.87 for the other three models [^1].

Background

Instead of generating an image, the model works like a person with pencils: it calls tools that set a color, lay down strokes, smudge or erase, and looks at its canvas to decide what to fix [^1]. With no solid fills, tone and color must be built up by layering [^1]. Runs with a target image were scored with SSIM, a structural similarity measure from 0 to 1 where higher means closer to the target [^1].

What happened

In a July 21 post, Hershal Bhatia of TryAI, a service that sells pay-as-you-go access to many AI models, described the company's "drawing arena" [^1]. It ran GPT-5.6 Sol, Claude Fable 5, Grok 4.5 and Gemini 3.6 Flash on two scored targets, the Mona Lisa and Van Gogh's Starry Night, and on five open-ended prompts such as a tabby cat asleep on a windowsill, for 28 drawings in total [^1]. Every model got the same toolset, and reasoning was set to high for all of them. The harness is open source [^1].

The models used the tools in very different ways. GPT-5.6 Sol never called the separate color, brush or pressure tools and set those values on each draw call instead. Of Grok 4.5's 1,349 tool calls, 65% went to those settings. Claude Fable 5 leaned on smudging, with 123 calls, and nearly a third of Gemini 3.6 Flash's calls were reviews of its own canvas [^1]. Claude Fable 5 averaged 12.5 minutes per drawing and an estimated $160.58 for its seven, against $7.74 for GPT-5.6 Sol, $9.21 for Grok 4.5 and $12.87 for Gemini 3.6 Flash [^1]. Grok used the most tokens, 34.0 million, but stayed cheap because about 98% of them were cached reads [^1].

On SSIM, Gemini 3.6 Flash scored highest on both targets. On the Mona Lisa it ended at 0.337 after a peak of 0.449, the best of the batch, ahead of GPT-5.6 Sol at 0.325, Claude Fable 5 at 0.286 and Grok 4.5 at 0.151 [^1]. The models also plateaued early. Claude Fable 5 reviewed its Mona Lisa 27 times, but its similarity was essentially flat after about the fifth review, and according to TryAI, in all eight scored runs the final drawing came in below the best the model had reached mid-run [^1]. The team's own verdict differed from the scores: it called GPT-5.6 Sol "the runaway leader", put Claude Fable 5 second on quality and wrote that Grok 4.5 "was basically garbage here" [^1].

Why it matters

This is a vendor's informal experiment, and the post says so plainly: "These are not objective tests of model capability" [^1]. Its value lies in what it shows about long, tool-driven tasks. The gap between the SSIM ranking and the team's judgment is a reminder that, as TryAI notes, SSIM measures structural closeness, not how good a drawing looks to a person [^1]. The review data also cut against the idea that more self-checking gives better results. Gemini 3.6 Flash reviewed its canvas most often, about 23 times per drawing, and still slid back from its peak, which led the post to conclude that the models "kept editing past their own best performance" [^1].

Cost is the other lesson. Claude Fable 5 took much longer and cost far more than the others for output the team ranked below GPT-5.6 Sol, although the same team says Fable has beaten GPT-5.6 on plenty of other tasks, including its music-video challenge [^1]. The post adds that the open-weight models it tried were not usable, with several returning a blank canvas [^1].

Practical example

Say you let an AI agent refine a chart or a slide layout by looking at its own render and fixing what it sees. TryAI's drawing test suggests two habits. First, keep every version the agent produces, because in the scored runs the final drawing was below the best intermediate one [^1]. Second, cap the number of review rounds, since Claude Fable 5's Mona Lisa barely changed after about the fifth of its 27 reviews [^1]. Then compare the result yourself instead of trusting a single similarity score, which here ranked the models differently than the people who looked at the drawings [^1].

Sources

  1. [1]TryAI — "Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok
  2. [2]Hacker News — "Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok