Same model. Same weights. Same GPU. Run it through two different harnesses and you'll get wildly different speed, quality, and cost. Nobody talks about this enough.
We recently watched deepseek-v4-flash-0731 hit 400+ tokens per second running inside an agentic coding harness — a 97% cache-hit rate, 3.3s average time-to-first-token, sustained across a 53-step, four-and-a-half-minute multi-turn session. That's not the model talking. That's the harness doing its job.
What a harness actually controls
A harness isn't just a prompt wrapper sitting in front of an API call. It's context management, tool-call structure, caching strategy, retry logic, and how multi-turn state gets batched across a session. Two harnesses calling the identical model, with the identical weights, can produce completely different effective throughput — because the model was never the bottleneck to begin with.
The 97% cache-hit rate is the real story here. That number reflects the harness structuring requests so the KV cache stays warm across the entire session instead of getting invalidated every turn. Miss that, and the same model looks three to five times slower for reasons that have nothing to do with the weights.
Benchmarks miss this entirely
This is why “which model should I use” is often the wrong question for anyone building an agent. The better question is which model-and-harness combination is actually running efficiently for the specific workload — and most public benchmarks test models in isolation, which means they miss the variable that's driving the biggest swings in real-world throughput.
As agentic coding tools mature, expect this to matter more, not less. The harness is becoming as much a performance lever as model choice itself — sometimes more. If you're evaluating inference speed by looking only at raw model benchmarks, you're missing half the picture.