OpenAI published first results for Jalapeño, its custom inference chip, on 25 August. Measured on InferenceX, a public benchmark from SemiAnalysis, against leading commercially available systems, Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency, rising to 2.1 to 4.1 times higher performance on highly interactive workloads. The tests ran across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T, so the architecture is not tuned to one in-house model.
The measurement method is the part that matters for agents. OpenAI evaluates at a matched user experience: how much useful work a system completes per unit of power while still meeting the latency interactive agents require. The company's stated reason is that agents run many steps in sequence, so per-step delay compounds across a task, which makes latency a throughput problem rather than a comfort metric.
That is the scheduling reality described in Gimlet treating inference as a multi-silicon problem and in inference sizing becoming explicit capacity planning, and it sits next to the control question raised by Anthropic's model hardware standard.
The numbers are OpenAI's own runs against unnamed comparison systems, and ramp plans are stated as intent. The condition to watch is whether independent InferenceX submissions reproduce the per-watt range, because a first-party benchmark result on a public benchmark is still a first-party result.