OpenAI says its own models now improve the chips that run them. In a research post on 18 November, the company estimated that design changes proposed by its models cut the standard-cell area (the silicon footprint of a block’s basic logic gates) by about 13% in the vector unit and 10% in the matrix unit of Jalapeño, its first custom inference chip. The numbers are OpenAI’s estimates, and no outside party has published a check.
The work began in early 2025 with GPT-o3 mini, a much weaker model than today’s. OpenAI says even that model found changes that mattered. The vector unit runs work in parallel, and OpenAI traces its savings to several arithmetic circuits: adders and multipliers for decimal numbers, format converters, reduction logic, and a unit for special functions. The models tried alternative logic layouts and different ways of splitting calculations across pipeline stages. That let designers share hardware and hold less temporary data.
The matrix unit did the heavy arithmetic. One multiplier redesign needed roughly 15% less cell area when synthesized, and it finished in one clock cycle instead of two. An adder redesign also saved roughly 15% of area while still taking two cycles. OpenAI says these changes were silicon proven during Jalapeño’s bring-up, meaning they worked on real chips and not only in simulation.
The same approach has moved to hardware OpenAI does not own. Working with AMD, OpenAI reports gains of up to 14% on power, performance, and area (the three standard scorecards for a chip, usually shortened to PPA) against heavily optimized baselines on AMD’s flagship GPU designs. The model got there by rewriting SystemVerilog RTL, the code that describes a chip’s circuits. OpenAI describes the result as roughly a year of progress in under five months.
According to OpenAI, AMD has already folded some of these changes into a design it recently sent to be manufactured, and it stored others in its production files for later versions. AMD separately verified PPA gains in the single digits across important blocks of its upcoming flagship CPU. That is the one claim in the post with a named outside party behind it, though the post does not quote AMD directly.
The model did not work unsupervised. Synopsys tools confirmed that each rewrite left the block’s behavior unchanged and scored what it cost or saved. Synopsys makes the electronic design automation (EDA) software that chip teams already use for exactly this checking, which is why OpenAI is now training GPT-Synopsys, a model built to drive those tools the way expert designers do.
To track progress across model generations, OpenAI tested five models on PicoRV32 and Ibex, two freely available processor cores built on the RISC-V instruction set. Each model was asked to raise throughput, the number of instructions a core completes per second. Any design that failed a simulation check or used more than 16 times the original area scored zero. Across 9,881 scored runs, GPT-6.2 Astra posted the best mean result on both cores when each attempt had roughly 112,000 tokens to work with. GPT-6.1 Sol beat GPT-5.6 Sol on both, and on PicoRV32 GPT-6.2 Astra scored about twice what GPT-6 Astra did.
Starting from a written specification beat tuning an existing design for all five models. GPT-6.2 Astra averaged about 0.7 on PicoRV32 and about 0.75 on Ibex that way. OpenAI itself flags the catch: freed from the originals, models could drop features like performance counters and extra instructions, which may do little for the test workloads. The company says stricter area caps and timing checks once the chip is physically laid out would bring the test closer to real chip work.
The unresolved question is how much of the 10% to 14% survives outside OpenAI’s own accounting. Synthesis-stage area numbers often shift after layout, and the RISC-V results stop short of that step. Chip teams weighing whether to put model-written RTL into a design flow should ask for post-layout figures before treating these gains as a budget line.
OpenAI, “Advancing chip design with OpenAI models,” published 18 November 2026 at openai.com.