OpenAI says GPT-Astra plus Codex wrote Jalapeño kernels 1.5 to 1.8x faster than human-expert code.

The claim covers selected attention and MoE blocks, alongside a claim that model-assisted kernel work brought three previously unplanned open-weight models to high performance on the new chip in about two months (Latent Space). Treat the numbers as first-party until the full Hot Chips presentation is out. The direction is the more useful signal: low-level performance work is now a plausible agent task rather than a place agents reliably produce plausible-looking garbage.