OpenAI says its models are now capable of improving real computer chips, including its own custom silicon and hardware made by AMD. In a new post, the company describes design changes its models suggested, and it introduces tests that measure how each new model generation handles chip design tasks.
The work builds on Jalapeño, OpenAI's first custom chip for running AI models. Engineers started using an early reasoning model, GPT-o3 mini, to help design it in early 2025. The team estimates that model-proposed changes could improve key arithmetic blocks. The changes were confirmed during the chip's silicon bring-up, meaning they worked in manufactured hardware, not just on paper.
From OpenAI's own chip to AMD and open-source tests
With AMD, OpenAI says its model rewrote SystemVerilog RTL, the code that describes a chip's circuits, and improved PPA, short for power, performance, and area, against already heavily optimized designs for AMD's flagship GPU. AMD has included some of these changes in a chip sent for manufacturing and verified smaller, single-digit percentage gains on key blocks of its next flagship CPU.
To track progress across model generations, OpenAI tested successive models on two open-source RISC-V processor cores, PicoRV32 and Ibex from the OpenTitan project. Models were asked to raise throughput, meaning instructions completed per second, without breaking correctness or growing the circuit too much. The newest model, GPT-6.2 Astra, scored best at comparable budgets. All five tested models did better when designing a core from written requirements than when editing an existing design.
This matters because chip design is slow, expensive, and done by scarce experts. If AI models can reliably find expert-level improvements, they could shorten design cycles and make AI hardware cheaper and faster, which would feed back into AI progress itself. OpenAI is also partnering with Synopsys, whose electronic design automation tools engineers already use to build and verify chips, to train a new model that uses those tools the way expert designers do.