Rails creator DHH just ran an AI coding model shootout, and the results are a little weird. He handed five models the same task — porting the Python library TerminalTextEffects to Rust — and let them rack up token bills. DeepSeek Pro V4 Max finished in about 2.5 hours with a total token cost of $23. Grok 4.6 spent $55 over roughly 1.5 hours, GPT-5.6 Sol burned $43, and two models never made it to the finish line: DeepSeek V4 Flash and GPT-5.6 Luna.
The earlier round belonged to Fable, which DHH estimates cost about $550 and took around 45 minutes. It rewrote TerminalTextEffects from Python to Rust using roughly 11 million tokens. The results are genuinely impressive: startup time dropped from 87ms to 2ms, and rendering got 9.6x faster.
But there's a big caveat. Fable first wrote a 483-line implementation plan, which Codex xhigh reviewed — and then every other model just followed that same plan. So this isn't really a test of each model's ability to solve a problem from scratch; it's a test of how efficiently different models execute an existing blueprint.
The project is open source. The repo includes 354 consistency tests that compare the Rust version against the original Python frame-by-frame and byte-by-byte, and the final product compiles down to a ~3.3MB static single file.
DHH shared the full results on X.