Most Agent Calls Don't Need Opus. LangChain Timed It.

LangChain pointed NVIDIA NeMo Switchyard at 145 multi-turn agent tasks and counted how many turns actually needed a frontier model. The answer was 7%. A 30B took the other 93%. The bill dropped 74%. Accuracy dropped 6 points.

That is the news. The rest is how I read it from a GX10.

What they actually measured

145 tasks, about 6.3 model calls each. Support dialogue, on-call investigation, multi-step workflow. The pairing is NVIDIA Nemotron 3.5 Lightning and Claude Opus 4.8. Escalation starts cheap. A judge watches. Two bad turns and that task moves to Opus for the rest of the session.

Numbers from the LangChain writeup, Aug 17:

7% of model calls went to Opus (range 4.1–9.1% across five runs). Those calls ate 68.4% of the spend. Lightning took 93% of calls for 10.4% of the spend. The judge is not free. It took 21.2% of routed spend and gets no prompt-cache gift.

Local AI Model Tracker (2026)

They also published the only formula I will steal:

minimum offload = judge cost / (expensive cost − cheap cost)

On their pairing they needed to offload 5.9% of turns. They offloaded 93%. Not close. The formula is for the cases that are close. Two APIs with a thin price gap cannot pay for a judge. Host the cheap model yourself and the gap opens again. They name an NVIDIA DGX Spark for that. I have the ASUS Ascent GX10 version of that box.

NVIDIA's own writeup adds a Cognition run on Devin Desktop: Opus 5 and Kimi K2.7, 50.6% at $3.11 mean, within 2.8 points of Opus 5, about 28% cheaper. Different bench. Different pair. Same claim class. I am not merging the two scoreboards.

How I read it from this box

I already run two lanes on one cube. Qwen 3.8 27B NVFP4 for chat. DeepSeek V4 Flash for the Moon Trail coding loop. I do not keep both resident. That is a human router with one seat.

Switchyard is the automated version of that habit. The news is not "routing is AGI." The news is a lab finally published the split: most agent turns are cheap work wearing a frontier invoice.

I have not installed Switchyard. The Deep Agents middleware is still experimental and not a package. Clone, local install, Python 3.12, deepagents 0.7.4+. Their headline numbers come from the server classifier in escalation mode, not from the in-process example. If you paste a tutorial and miss that, you will not reproduce 7%.

What I will not do is put 27B and Flash on the same GX10 to cosplay this paper. One heavy stack at a time. Same rule as the Qwen receipts.

When this is not the story

• Latency. The judge is a second call, about 700ms. The stage router is the cheap-latency path. They did not bench that one.
• Short chats. Escalation needs a trajectory.
• Saturated evals. They say the suite only had 8 points between a 30B and Opus. Harder work would move the 6-point hole. Treat it as one workload.

The 2026 version of pick a lane

Cloud Grok and Cursor Ultra stay the paid seats when the task is actually hard. The GX10 stays the cheap seat. A router is interesting when you cannot tell which request just arrived.

Until I run it here, this is a quote of their table, not a WikiWayne bench. Follow the corrections on X (@wikiwayne).