Agents Beat Two Accelerator Chips on Inference in Under 24 Hours, No Human in the Loop

Here is where one agent-tuned B200 lands against the fastest gpt-oss-120b providers Artificial Analysis tracks, and how it got there overnight.

Faster than two accelerator chips

On Artificial Analysis, the fastest gpt-oss-120b providers aren't GPU clouds. They are companies that built their own inference accelerators: Cerebras (wafer-scale engine), SambaNova (RDU) and Groq (LPU).

One general-purpose GPU, tuned overnight by agents, now runs ahead of SambaNova (706.8) and 53% faster than Groq (473.5), and about 2× the other GPU providers on the list.

These are chips designed from the ground up to make exactly this workload fast. The agents closed that gap with software alone, on a GPU you can rent today, in a single night.

1.5× faster, overnight, on its own

At noon UTC on October 15 we gave four agents one NVIDIA B200, the open gpt-oss-120b weights and one goal: make a single request stream as fast as possible. No human touched the engine after that.

They started from a strong, standard recipe at about 480 tokens per second. Twenty hours later, the same GPU streams 723.2 tokens per second: 1.5× faster, found, tested and verified entirely by agents.

How the agents did it

Four agents ran the loop described in Recursive Self Improving Inference Engine: propose a change, submit it to the harness, keep it only if it beats the current best, and share what they learned so the others can build on it. Over the night they ran 83 scored experiments and exchanged 501 messages.

The biggest jumps landed around the eight-hour mark, as the agents' early work started paying off. After that came a steady run of smaller wins, each one checked against a freshly re-measured control before it was kept. Ideas that didn't hold up were logged and rolled back, so every gain on the progress chart is a verified result.

How it's scored

Every submission is measured by a fixed harness the agents can't modify, on the Artificial Analysis performance methodology: one request at a time, 10,000-token input prompts, and output speed counted from the first token to the last. A result only counts if it also keeps time to first token at or under 1.5 seconds and errors at or under 1%.

The current best reaches first token in 108 ms on a 10k-token prompt, generates a median 1,491 tokens per answer, and has zero failed requests.

What's next

The line is still rising. The agents' next milestone is 1,000 tokens per second on one GPU, and the target after that is Cerebras at 1,729.0 tokens per second.

The same loop works on any open model and any traffic shape. Bring the workload, the speed you need and the quality bar it must hold, and the agents do the tuning.

Sources

Artificial Analysis, gpt-oss-120b (high) providers: artificialanalysis.ai/models/gpt-oss-120b/providers, accessed October 16, 2026. Output speed is the P50 over the past 72 hours on 10k-token inputs.