Twenty model calls turn a two-second step into a 45-second wait

A developer waiting on a coding agent is waiting on every call the loop makes. Red Hat's post splits a 20-call task two ways: at roughly 2 seconds a call it runs to something like 45 seconds of wall-clock time, and with responses coming back in well under a second the same task finishes near 13 ^[4][5]. Per call that is 2.25 seconds on the slow path and 0.65 on the fast, and the fast backend gets through the loop about 3.5 times sooner ^[1][2][3]. Those are the post's illustrative figures; it does not report timings from a deployed agent.

Both numbers assume short outputs. The post describes each step as another few thousand tokens of reasoning before the agent can act on what it learned ^[2]. The low end of that is 2,000 tokens, against the roughly 40 tokens per second Artificial Analysis records for Kimi K3 and flags as slow ^[8]. One step is 50 seconds of generation, and twenty of them is about 16 minutes ^[4]. Qwen3.8 Max sits in similar territory at around 47 tokens per second ^[9].

Teams tell themselves they are picking the model that gets the hardest step right. The person at the keyboard sits through a queue of dependent calls where nothing overlaps, because the agent cannot start step four until step three returns ^[3]. Red Hat's author says the loop was "painfully slow" and that a developer waiting on a coding agent has an attention threshold, and that once it is crossed the tool stops feeling like a collaborator ^[1][14]. That threshold stays unquantified. It cites Andrej Karpathy's 2025 point that as models get more capable the human's job moves toward orchestrating and reviewing agent work, and that slow or idle agents make the human the bottleneck ^[15].

The price of accuracy has been measured. Epoch AI studied accuracy against runtime across major providers and found that halving a model's error rate tends to slow it by roughly 2 to 6 times, about 6 on GPQA Diamond and closer to 2 on a competition math benchmark ^[6]. Apply that to the 13-second loop: half the errors costs somewhere between 26 and 78 seconds ^[5]. Epoch also found the efficiency frontier populated overwhelmingly by smaller distilled variants, the ones with "flash," "mini," or "nano" in their names ^[7].

So the forcing function is two numbers, and only one of them comes from a vendor. The wall-clock ceiling a team will accept for a loop, divided by the calls that loop makes, sets the per-call budget. A 20-second ceiling over 20 calls leaves 1 second a call. At 40 tokens per second, that second buys about 40 tokens of output ^[6]. Candidates then get checked against the time-per-task and cost-per-task axes Artificial Analysis publishes, alongside the intelligence index ^[13].

A model that fails the per-call budget still has a job. It belongs on the single planning call at the top of the run, where one long think is affordable, and off the inner loop. Nvidia has already positioned Nemotron 3.5 Lightning for the other end, trading some intelligence for speed as a fast execution layer ^[10]. Red Hat puts Gemini 3.7 Flash and StepFun's Step 3.7 Flash on the speed-versus-intelligence frontier, quick enough for tight loops while still scoring relatively well for their tier ^[11].