OpenAI published the first performance results for Jalapeño, its first custom-built inference chip, and the numbers show a working piece of silicon rather than a roadmap slide. Three public models carried the test load: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Peak-load throughput per watt landed between 1.5 and 1.9 times higher than the comparison hardware, while end-to-end latency came in 1.7 to 3.6 times lower. Push the workload toward highly interactive use and the gap stretched further, to 2.1 to 4.1 times.
The stakes are straightforward. Inference, not training, is where most AI compute spend now goes, and the chip that serves an agentic workflow faster and cheaper per watt changes the unit economics of every product built on it. OpenAI framed the release as proof that a company designing its own models, serving software, and silicon together can move each layer faster than a company buying commodity accelerators alone.
OpenAI ran the tests on InferenceX, a public benchmark built by SemiAnalysis that measures the full request-serving pipeline rather than raw chip throughput. That is a meaningful choice: it lets outside observers reproduce the comparison rather than take OpenAI’s word for a proprietary number. Even so, OpenAI selected the operating range and the comparison hardware itself, and the results have not been reproduced by an independent lab. A public benchmark is not the same thing as an independently verified test run.
Power efficiency is where the chip’s design choices show up most clearly. OpenAI rates the chip at 700 watts, yet measured sustained draw stayed at 550 watts or under across every workload tested, a gap the company did not fully explain. The comparison hardware ran hotter throughout: Nvidia’s GB200, rated at 1,200 watts, served as the baseline for GPT-OSS 120B, while the 1,400-watt GB300 stood in for both DeepSeek R1 and Kimi K2.5. Kimi K2.5, the largest of the three models, produced the tightest published margin: roughly 1.5 times the peak performance per watt and a 3.4-times cut to latency. OpenAI also claimed a wider advantage on its own frontier models internally, though it published no numbers to support that claim.
The chip’s origin story is arguably the bigger news than the benchmark charts. Design to tapeout took nine months, a timeline OpenAI credits partly to its own models: earlier generations helped with chip design and bring-up, and newer ones are now speeding up optimization and programming work. Two months is how long it took the team, pairing Codex with an internal model called GPT-Astra, to bring three open-weight models to high performance that had not been part of the chip’s original production plan. One narrower result stands out inside that broader claim: for select attention and mixture-of-experts kernels in GPT-OSS, code generated by AI ran 1.5 to 1.8 times faster than the human-written version it replaced. OpenAI limited that comparison to those specific kernels rather than the model as a whole.
OpenAI describes this as only the opening generation of a multigenerational platform, with a second generation already deep in development and a third taking early shape. Production deployment inside OpenAI’s own infrastructure is targeted before the year is out, and that rollout is the point at which these benchmark numbers get tested against actual customer traffic instead of a controlled harness.
For any team building latency-sensitive agents on OpenAI’s API, the practical question for the next ninety days is whether Jalapeño’s deployment changes pricing or rate limits before Gen 2 ships, since a chip that serves more tokens per watt gives OpenAI room to cut inference costs without cutting margin.
Reported by OpenAI in a first-party technical post published on openai.com.
