GPT-5.5 at its xhigh reasoning setting took the top of KernelBench at 6.57%, per Greg Brockman, while OpenAI’s Codex engineering lead Thibault Sottiaux wrote that Codex had achieved escape velocity and would keep improving rapidly (X). The kernel result is the first kernel-writing leaderboard in the corpus, making concrete the early sign Jack Clark named in January and following Opus 4.6’s 34x training speedup and the twelve-hour CPU in the silicon-designs-itself thread. Sottiaux’s remark updates his own February line that Codex builds itself. The storyline continues with agents writing AMD kernels, a model rewriting the kernels behind an 80% price cut and, in August, the loop reaching silicon.
Development
A model tops the leaderboard for writing the GPU kernels it runs on
The newsletter reads a kernel-writing leaderboard and a lead engineer's 'escape velocity' remark in one breath, as the self-improvement loop baked into the dev cycle and now reaching down into the hardware layer.
Claim
GPT-5.5 topped KernelBench at 6.57% for writing GPU kernels, meaning the model is now optimizing the hardware that runs it, while OpenAI's Codex lead declared the product has achieved escape velocity and will keep improving rapidly.
Reported metric: 6.57%
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to April 29, 2026 (theinnermostloop.substack.com)
- Greg Brockman on X: GPT-5.5 xhigh tops KernelBench (x.com)
- Thibault Sottiaux on X: Codex has achieved escape velocity (x.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
Related to
Reported in
Evidence for
Supersedes
Referenced by
- The speedrun record falls to 106.9 seconds on compiler kernel hacking
- A model runs thirty-five hours without a human
- A topological quantum chip is designed with the company's own agent
- A model speeds up training code thirteen times better than a skilled human
- Seventeen models have taken the lead, each reigning about seven weeks
- A challenger GPU serves an open model at half the cost as agents write kernels
- A model autonomously rewrites the kernels behind an 80% price cut
- Self-improving agents rebuild the inference stack serving two open models
- An agent architecture lifts a model from 30% to a perfect 100
- A model running on the chip finds optimizations for its own operations
- Welcome to April 29, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.