@dhh
Post
GPT-5.6 Sol High was able to follow Fable's plan and produce a great version too. First stab was 30% slower, but one follow-up prompt got it to parity. Per-token cost: $43.
In contrast, DeepSeek V4 Flash couldn't get anything working despite many follow-up prompts.
Frontier!
Image from X post
Quoted post by DHH (@dhh) Fable one-shotted a Rust rewrite of the TerminalTextEffects Python library in 11M tokens. Startup time went from 87ms to 2ms and rendering speed is up by 9.6x. Now zero dependencies and a 3mb single exec 🤯 https://claude.ai/code/artifact/287825bc-7aad-4541-a7e7-fa4ba8d03612
Explanation
What it says DHH reports that GPT-5.6 Sol High successfully implemented Fable’s plan for a Rust rewrite of the Python `TerminalTextEffects` library. Its first attempt was ~30% slower than Fable’s version, but one follow-up prompt brought it to performance parity. Total model cost was about $43. By contrast, DeepSeek V4 Flash failed to produce a working implementation even after many follow-ups.
The quoted earlier post says Fable one-shotted the rewrite using 11M tokens, reducing startup from 87 ms → 2 ms, increasing rendering speed 9.6×, eliminating dependencies, and producing a 3 MB standalone executable.
Context This is less a raw model benchmark than an agentic coding comparison on a concrete, unusually large task: follow an existing architecture/plan, implement it in Rust, then optimize until it matches a strong reference result.
Why it matters The interesting signal is that frontier models may now be capable of executing substantial software rewrites economically, while model quality still varies sharply: one model converged with a single correction; another apparently never reached working code.
Images The $42.95 bill was dominated by 59.46M cached-input tokens ($29.73) and 251k output tokens ($7.54); fresh input cost only $5.68.