OpenAI published the first performance results for Jalapeño, its custom AI inference chip, on August 25, 2026. Across three public models, OpenAI reports 1.5–1.9 times more AI work per watt at peak throughput and 1.7–3.6 times lower end-to-end latency than the comparison systems.

The company says Jalapeño is designed for interactive and agentic workloads, where delays compound across many sequential model calls. It plans to begin deploying the chip inside OpenAI’s compute infrastructure by the end of 2026; this is a performance announcement, not a claim that every ChatGPT, Codex or API request already runs on Jalapeño.

What OpenAI measured

  • GPT-OSS 120B: about 1.9× higher peak mixed throughput per kilowatt and 1.7× lower end-to-end latency in the published comparison.
  • DeepSeek R1 670B: about 1.7× higher peak performance per watt and 3.6× lower end-to-end latency.
  • Kimi K2.5 1T: about 1.5× higher peak performance per watt and 3.4× lower end-to-end latency.

OpenAI used the public InferenceX benchmark and says it compared systems at a matched user experience. The figures are workload- and configuration-specific, so they should not be treated as universal speedups for every model or application.

Why this matters for QA engineers

OpenAI Jalapeño agent latency matters because faster inference changes observable product behavior, not only infrastructure cost. Existing test thresholds may become too loose, and concurrency can expose bottlenecks that were previously hidden behind model response time.

  • Re-baseline end-to-end SLAs: measure time to first token, inter-token delay, full response time and total agent-task duration separately.
  • Keep timeout tests realistic: do not simply shorten every timeout. Validate cold starts, long context, tool calls, retries and peak-load tails before changing failure thresholds.
  • Stress downstream systems: faster model output can increase pressure on browsers, APIs, databases, queues and rate limiters used by an agent.
  • Test streaming assumptions: confirm the UI renders rapid token bursts correctly without dropped events, duplicated text or broken cancellation.
  • Track quality with speed: compare task success, tool-call accuracy and recovery behavior at the same time as latency. A faster wrong result is still a failed release.

A practical performance test

Build one repeatable agent scenario with five to ten sequential steps, including a tool call and a retry. Record p50, p95 and p99 latency for each model call, the complete task, and every downstream dependency. Run it at low and peak concurrency, then compare both speed and task success after any serving-infrastructure change.

This separates genuine inference improvement from queueing, network or tool latency—and prevents a faster model layer from masking a slower overall workflow.

Bottom line

Jalapeño’s first results suggest lower latency could make long, multi-step AI agents substantially more responsive. For QA teams, the near-term action is preparation: preserve current baselines, instrument every stage of the agent loop and test how the rest of the system behaves when inference stops being the dominant delay.

Sources