DHH had Opus 5.5 rewrite a Rust screensaver in assembler
The Omarchy ttfx engine went from Rust to hand level SIMD in a single translation, and got faster while he slept.
DHH says he ported the Omarchy screensaver engine, ttfx, from Rust to x86-64 assembler and it came out up to 17x faster. He describes it as a one-shot translation by Opus 5.5.
We keep drilling until the agentic drill bit hits bedrock!
Then it kept going. Overnight, DHH says Opus improved efficiency further, to up to 22x faster at peak and 6x faster on the mean, and up to 450x faster than the original Python implementation. The assembler version supports SSE2, AVX2 and AVX-512, and falls back to the slower Rust implementation otherwise.
While I was sleeping, Opus improved efficiency further.
Why this one is different
Most of the agent results people post are feature work, refactors and test suites, the sort of thing a competent engineer could have done with more hours. Hand written SIMD assembler with three instruction set paths and a safe fallback is not that. It is the category of work most teams skip entirely, not because it would not pay off but because nobody on the team writes AVX-512 by hand and nobody wants to own it afterwards.
That also changes what the numbers mean. A speedup against a Rust baseline that was itself a port of Python is a speedup against whatever the original author happened to write, not against a tuned reference. DHH gives peak and mean figures for the assembler version and a separate comparison against the Python original, and he does not publish the benchmark methodology in these posts. Take it as one engineer's result on one small graphics engine, which is what it is.
The drill bit line
DHH's framing is that the agents keep finding new floor. "We keep drilling until the agentic drill bit hits bedrock" is how he put it, which is a more useful way to read this than as a benchmark claim. The interesting part is not that a model can emit assembler, it is that the output was apparently good enough to ship behind a runtime fallback and then get iterated on unattended.
For working engineers, the fallback is the load bearing detail. Keeping the Rust implementation as the path for machines without the right instruction set support means the assembler is an optimization you can switch off, not a rewrite you have to trust. That is roughly the same containment strategy other people have been applying to agent output, from Gary Bernhardt cutting diffs down before review to Mitchell Hashimoto's rule that you should be able to explain the code before shipping it. Nobody is claiming they read the AVX-512 line by line.