TryAI put four frontier models in front of a blank canvas and a set of virtual colored pencils. GPT-5.6 Sol, Claude Fable 5, Grok 4.5 and Gemini 3.6 Flash were tasked with reproducing the Mona Lisa and other images using the same tools. Across seven drawings, Claude Fable 5 spent $160 in tokens versus $7.74 for GPT-5.6, producing worse results on this specific task (TryAI data, July 21, 2026). Fable used 14.6 million tokens in an average of 54 steps, GPT-5.6 used 3.4 million in 29 steps.

That same day, the Claude Code team shared in a public interview how Anthropic uses its own tools. The Claude Code system prompt was cut by 80 percent: recent models no longer need examples in the prompt, and prohibition lists actually degrade quality. Claude Tag, the collaborative Slack integration, closes 65 percent of the engineering team’s PRs.

Standard benchmarks tell you which model is most capable on average. TryAI’s test tells you how much it costs you to produce usable output on a concrete task, and shows that the price-to-quality ratio flips completely from one task to another. Claude Fable 5 beats GPT-5.6 on other tasks, as we reported in the music clip challenge, but here it costs twenty times more for a worse result.

Operational economics is the data you need when choosing a model for your work, not the overall ranking. Anthropic itself works with this logic: it reduces the prompt and automates PRs, testing features on its own employees before release.

If you want to measure real cost on your own cases, the playbook for comparing two models in fifteen minutes gives you the method.

In detail

TryAI’s test is built to be readable at a glance. Each model receives the same set of virtual tools: choose color, stroke thickness and pressure, draw batches of marks, blend, erase, and call view_canvas to see your work and correct it. No solid fills: tone is built in layers, like with real pencils. Similarity to the target image is measured with SSIM, an objective metric of structural similarity.

The numbers tell a clear story. Claude Fable 5 used 14.6 million tokens across seven drawings, versus 3.4 million for GPT-5.6. It averaged 54 steps versus 29, taking 12.5 minutes versus 6.2. And it spent $160 versus $7.74. The visual result, measured by SSIM, was worse.

This means Fable is a model with weak points on this type of specific task, even though it excels in others. TryAI flags this openly: Fable had won against GPT-5.6 in the previous music video challenge, and Anthropic uses it internally for tasks where it excels. On this type of task, a model that reasons heavily and does many self-revisions burns budget without proportional returns. Grok 4.5 used 99 steps and 354 draw calls, but started poorly and struggled to correct itself. Gemini 3.6 Flash spent more time looking at the canvas than drawing.

The four models used the same tools in completely different ways. GPT-5.6 never called set_color, set_brush or set_pressure separately: it set everything inline in every draw call. Claude Fable averaged 15.6 self-revisions per drawing, double anyone else.

The interview with the Claude Code team tells the same economics from the tool-builder’s side. Cat Wu and Thariq Shihipar described a shift in approach: with Fable you can complete in a single attempt features that previously required constant oversight. The Claude Code system prompt was cut 80 percent because recent models no longer need examples in the prompt. Lists of “don’t do X” reduce result quality with new models.

Thariq said something that sounds heretical for a software engineer: rewrites now work. If you have a good test suite, rewriting from scratch with an agent costs less than maintaining legacy code. The code itself is a specification, and often the only copy you have.

TryAI’s test limitations are stated upfront. The tests are intentionally open and fuzzy tasks. The open-weight models tested weren’t even usable, several returned a blank canvas. The test separates frontier models from the rest but doesn’t say which is better overall. It says how much it costs to produce output on that task, and whether it’s worth what you spent.