Opus 4.7 is lazy

On OpenRouter, with reasoning high and max_tokens 100k settings, Qwen 3.6, Opus 4.6, and Opus 4.7 were compared on the same pelican-bicycle SVG prompt.

• With the basic prompt, Opus 4.7 produced the shortest output, using the least tokens at around 20 tokens.
• Even with the think hard prompt added, 4.7 stayed around 95 tokens, while 4.6 and Qwen each used about 500 tokens.
• When structured instructions and a viewBox="0 0 1200 1200" condition were added, all three models produced plausible SVGs, but 4.7 still produced the smallest result.
• Total costs were Qwen 3.6 $0.0653 / 40.7KB, Opus 4.6 $4.0250 / 34.5KB, and Opus 4.7 $0.6447 / 22.6KB.

4.7 was much cheaper than 4.6 and also produced smaller results, but the experiment was a SVG generation benchmark run only once each, so there are limits to generalizing from it.