One model vs automatic routing: what our enterprise test shows

On NuRoute's published 426-prompt enterprise test, Artemis at Balanced received a mean judge grade of 0.952. A fixed Claude Sonnet 5 route received 0.949; a fixed gpt-5.5 route received 0.977. The reported costs per 100,000 requests were $141, $3,988 and $5,596 respectively.

Those figures make routing worth investigating. They do not show that every business application will get the same savings, or that a 0.003 grade difference establishes one system's superiority.

This article interprets NuRoute's published benchmark documentation, checked on 29 October 2026. We have not rerun the experiment for this article. All benchmark figures below come from that source.

What was compared

The documented test covers 426 held-out business prompts across 15 industries. It includes tasks such as extraction, summarisation, classification, drafting, analysis and coding. Single-model baselines answer every prompt; each Artemis preset is evaluated on its selected model.

These are benchmark accounting figures, not an all-inclusive NuRoute invoice forecast. Use the production billing rules and your actual token volumes to estimate direct service charges.

The inexpensive baseline matters

The comparison with gpt-5.5 shows a large cost gap. The comparison with DeepSeek-V4-Flash asks a more demanding purchasing question: how much does selection add over an inexpensive fixed model?

On this test, Balanced's reported cost is about 8.5% higher than that fixed baseline, with a mean grade that is 0.009 higher. Whether that difference justifies routing depends on the failures behind the averages, the uncertainty and your workload. The mean alone is not enough.

This is why we recommend testing a capable inexpensive model alongside Artemis. Choosing only an expensive baseline makes cost reductions easier to demonstrate without proving a router is necessary.

Why this test favors Artemis

The documentation explicitly states that the test is a good fit for the system. Training and test prompts come from the same broader corpus, with similar industries, task types and writing styles. Workloads were held out, but the kind of work is familiar.

The tasks also leave room for inexpensive models to perform well. That creates an opportunity for large savings against a frontier model used for every request. Harder or more unusual workloads can produce a different tradeoff.

The industry table is a different analysis: it combines 426 held-out prompts with 1,049 training prompts scored through workload-level cross-validation. Its 1,475-task total should not be described as 1,475 untouched test prompts.

How to check three workloads in your own product

For extraction, compare exact values, missing fields and invented information. For summaries, identify facts that must survive and unsupported claims that must not appear. For drafting, use a rubric covering required content, factual grounding and tone.

Keep those scorecards separate. A good result on routine extraction should not offset a failure to preserve a deadline in customer-facing summaries. Publish a weighted aggregate only after reporting the task breakdown.

The published enterprise table does not separate these three task types. The scorecards above describe tests to run on your own product; they are not additional NuRoute results.

Separate other benchmarks from this result

NuRoute's documentation also reports RouterArena and exam results. They use different datasets and scoring rules. The exam charts include calibration questions; the RouterArena result is reported by NuRoute, with listing approval described as pending in the documentation. Do not label all of those numbers as independently verified held-out results.

For a purchase decision, replay your own work. Use the evaluation guide to compare a fixed model and Artemis with the same acceptance rules, and inspect where routing is unlikely to help.