How it works · Updated 28 October 2026
How a NuRoute answer gets made
Every request goes to one model, chosen for it by Artemis, NuRoute’s model-selection engine. This page covers the prices, the benchmark results and how the choice is made.
Published rates
NuRoute Chat
A month, VAT included, with a weekly token allowance.
- Lite
- $3.99 · 5M tokens a week
- Personal
- $10 · 50M tokens a week, with API
- Personal Max
- $100 · 600M tokens a week, with API
Artemis on the Platform API
Per token, from prepaid credits, whichever model Artemis picks.
- Input
- $0.75/Mtok
- Cached input
- $0.03/Mtok
- Output
- $3.20/Mtok
Name a model instead and you pay its provider’s list price with no fee, or nothing to us on your own provider key. Pricing
Benchmark scores
Public tests, each against a named frontier model answering every question. Shown at Balanced, the preset NuRoute Chat runs.
- NuRoute, Balanced
- 88.9%
- gpt-5.5
- 92.6%
- NuRoute’s cost
- 2.4%
190 graduate-level science questions. Cost is NuRoute’s as a share of gpt-5.5’s. Hover a point for its numbers.
| Setup | Correct | Per 1,000 |
|---|---|---|
| gemini-3.1-pro | 93.2% | $79 |
| gpt-5.5 | 92.6% | $57 |
| Claude Sonnet 5.5 | 90.5% | $9.31 |
| Artemis, Quality-first | 89.5% | $19 |
| Claude Opus 5.5 | 89.5% | $19 |
| Artemis, Balanced | 88.9% | $1.37 |
| gpt-6-sol | 88.4% | $5.95 |
| Kimi-K2.6 | 88.4% | $82 |
| Artemis, Cost-first | 87.4% | $0.84 |
| gpt-5.6-luna | 87.4% | $1.63 |
| gpt-6-luna | 86.8% | $0.80 |
| Claude Haiku 5.5 | 84.2% | $0.95 |
| Qwen3-235B | 76.3% | $2.80 |
| Nemotron-3.5-Lightning | 76.3% | $3.83 |
| ministral-8b | 53.7% | $0.32 |
- NuRoute, Balanced
- 34.0%
- gpt-5.5
- 40.8%
- NuRoute’s cost
- 3.3%
500 text-only questions. Cost is NuRoute’s as a share of gpt-5.5’s. Hover a point for its numbers.
| Setup | Correct | Per 1,000 |
|---|---|---|
| Claude Opus 5.5 | 56.6% | $64 |
| Artemis, Quality-first | 54.4% | $62 |
| Claude Sonnet 5.5 | 46.4% | $51 |
| gemini-3.1-pro | 45.8% | $202 |
| gpt-6-sol | 41.8% | $17 |
| gpt-5.5 | 40.8% | $139 |
| Kimi-K2.6 | 35.0% | $162 |
| Artemis, Balanced | 34.0% | $4.61 |
| Claude Haiku 5.5 | 31.0% | $3.62 |
| Artemis, Cost-first | 29.6% | $2.46 |
| gpt-6-luna | 28.0% | $1.60 |
| gpt-5.6-luna | 23.8% | $2.82 |
| Qwen3-235B | 12.0% | $3.17 |
| Nemotron-3.5-Lightning | 11.2% | $6.86 |
| ministral-8b | 4.2% | $0.50 |
| Benchmark | NuRoute | Frontier | Cost |
|---|---|---|---|
| RouterArena, vs GPT-5 | 75.5% | 74.0% | 6% |
| GPQA Diamond, vs gpt-5.5 | 88.9% | 92.6% | 2.4% |
| Humanity’s Last Exam, vs gpt-5.5 | 34.0% | 40.8% | 3.3% |
Frontier: the named model answering every question. Cost: NuRoute’s cost as a share of it. RouterArena is independent.
The mechanism
Artemis predicts which model will give the best answer for the money and sends the request there. Answers come from one model and are never blended.
Your request
From chat, your tools or your code
One model name
artemisin the Chat API,autoin the Platform APIArtemis picks
Scores every model in the pool, across providers, and chooses one
One answer
From that model. Nothing is blended.
1
Read the request
Artemis works out what kind of task the request is, such as extraction, summarising, coding, translation or reasoning, or a mix of them.
2
Expect a score per model
For every model in the pool, measured scores on those task types give the expected quality of its answer. The scores come from graded answers, not from vendors.
3
Expect a cost per model
Artemis estimates how long each model’s answer will be, and prices the request and that answer at the catalogue price.
4
Apply the preset
The preset turns scores and costs into one choice: the best expected quality after a penalty for extra cost. Your cost and latency caps rule models out first.
5
Send it, with a fallback
The request goes to the chosen model. If that model fails, the next best one takes over.
Choosing adds no extra AI call and no tokens to your bill. Prompt text is redacted before Artemis reads it, and Artemis makes no network calls.
In a long conversation, Artemis can keep the model that answered the last turn when the provider’s cache makes that cheaper (rolling out). If Artemis can’t decide, a simpler fallback router does, and if a provider fails, the next one takes over.
Presets set the trade-off
NuRoute Chat runs on Balanced. On the Platform API you choose a preset, or name a model and skip the choice.
| Preset | Where | How it picks |
|---|---|---|
| Balanced | Default, and NuRoute Chat | Best quality after a moderate penalty for cost |
| Cost-first | Platform API | Best quality after a heavy penalty for cost |
| Fastest | Platform API | Balanced, with the time to answer counted too |
Estimating your bill
In NuRoute Chat, your plan’s weekly allowance is the limit. The sidebar shows what you have used and when it resets.
On the Platform API, multiply your input, cached and output tokens by the rates above. A typical request of 2,200 input and 500 output tokens costs about $0.0033 with Artemis, against about $0.056 on gpt-5.5.
What you see
- In NuRoute Chat
- The answer. Chat doesn’t show which model wrote it.
- On the Platform API
- Every response names the model that served it, and the dashboard keeps a trace of each decision, with the candidates it weighed. Routing reference
Check it yourself
Read the method
Every setup, every score and how the tests were run.
Run your own prompts
Send work you know the answers to and compare quality and cost with your current model. New accounts get $10 of credits.
Try the chat
Ask it the hardest question you asked another assistant this week.
Talk to us
Walk through your workload with the people who built Artemis.