Smart Router

The Pentos Smart Router automatically picks the right AI model for your message, so you never have to think about it. It analyzes your prompt and conversation context using pattern matching techniques, all processed securely on our servers in milliseconds with no impact on your response time.

It considers both the complexity and the type of task, working out how capable a model your message actually needs, then choosing the most efficient model that clears that bar. It recognizes what you are asking however you phrase it, so a math or coding question reaches a model that is genuinely strong at that work whether you write it formally or in plain language, while everyday conversation stays on fast, lightweight models. The result is faster responses on everyday work, at lower cost and lower estimated energy than routing everything to a flagship model, all without any extra effort on your part.

How It Works

Every Smart Router decision flows through five steps, calculated on our servers in milliseconds without calling another AI model:

1. Prompt analysis. We extract signals from your message: code, math, creative writing, translation, summarization, attachments, ambiguity, multi-part requests, and more.
2. Conversation context. We look at the rest of the chat to detect follow-ups, ongoing complexity, and the task type of the previous response so the AI keeps a consistent thread.
3. Difficulty score. Each detected signal contributes points toward a 0 to 100 difficulty score. This reflects not just your prompt but how much work it implies: the effort level you have set, whether tools or connectors are enabled, the size of the conversation, and any attachments all factor in. So even a simple-sounding message can score higher when it carries real load.
4. Capability bar. The score sets a minimum capability score for the category that matters for your task, using independent scores from BenchLM: coding requests are measured on coding, research and tool-heavy requests on agentic work, math and reasoning on knowledge, and everyday chat on overall capability.
5. Most efficient and capable model. The router then picks the most efficient model that clears the bar. When two models cost exactly the same, the stronger one wins. Web search and connectors are only used with models whose scores show they handle them well. If your message includes an image, the router only considers models that can see images; if it includes a PDF, it only considers models that read PDFs directly. It checks both the file you just attached and any files earlier in the conversation.

This is the “cheapest model that does the job” principle applied literally: a hard task raises the bar and the router pays for the capability it needs, while an easy task lets an efficient model win no matter how the request is phrased. Attaching a file or two does not push you to a bigger, costlier model on its own. Larger or more numerous attachments, or asking it to compare and combine information across several files, are what raise the score. The actual size of an attached file, not just how many you send, is counted toward picking a model with a large enough context window.

Difficulty Bands and Price Classes

The difficulty score maps to one of five bands, and each band sets how high the capability bar goes. Separately, every model belongs to a price class based on its output cost per million tokens. The two are related but independent on purpose: a demanding message raises the bar, and if an inexpensive model genuinely clears it, that model still wins.

Models above $15 per million output tokens are outside the Smart Router rotation entirely. They are available whenever you select them yourself from the model selector. Claude Opus 5.5 uses points like every other model; Claude Fable 5.1 and GPT 6 Astra run only on your own API key.

A response costs points by the model’s output price, not its price class: see How Points Work.

Backup Models

Along with its pick, the router lines up two backup models from other providers. If the chosen model’s provider is unavailable before it starts responding, Pentos switches to a backup automatically and lets you know. Your points are charged for the model that actually answered, and Message Info shows it.

Benchmark Scores

Model capability scores come from BenchLM, which combines public benchmark results into overall, agentic, coding, and knowledge scores. We record each model’s scores as BenchLM publishes them and update them as new models ship, so the router keeps surfacing the right model for each task. Where BenchLM has not yet published a score for a model, we may use another source’s figure or an estimate until it does. A model with no score stays available when you pick it yourself, but the router does not choose it automatically.

Effort Levels

The effort level you choose now steers routing directly. Higher effort raises the difficulty score, and picking a specific level (anything other than Auto) tells the router to choose a model that actually supports that level of reasoning effort. Auto lets Pentos pick a sensible effort for the selected model after routing. If no efficient model supports your chosen level under your current context mode, the router falls back gracefully and records that in the message info. Choosing High or Max also adds points to each message, and Auto never does. Max needs a paid plan: on Free, choosing it opens the upgrade options.

Energy Savings

Routing to the right model is also where the energy savings come from. When a lighter model can handle your message, it uses far less energy than a top-tier one. Routing a task to a Lite-class model instead of a flagship can drastically lower its estimated energy, and a Nano-class model often saves more still. The savings apply only when a lighter model is genuinely up to the task; when your message needs more power, the router escalates on purpose. See Energy Usage for worked examples and the full breakdown.

Pairing With Context Modes

Smart Router and Context Modes are designed to work together, and Lean Mode now shapes routing directly: with Lean Mode on, the Smart Router only considers models up to the Standard price class ($5.00 per 1M output), reaching for Nano and Lite models wherever they are up to the task. Standard and Max context modes let the router use the full rotation, up to the Pro price class. Lean Mode paired with Smart Router keeps both the model and the context small, which is where the cost and energy savings come from. We recommend leaving both on as your default.

Model Access

Every plan routes across the same full model lineup. The Smart Router treats all models equally and simply picks the most efficient one that meets your message’s capability bar and costs no more points than you have left, once the points your chat settings add are counted. When your points run out, it uses models that cost 0 points, which work in Lean Mode with Web Search and Connectors off. Models in the Max price class (over $15 per 1M output) are never routed, so you select them yourself.

Tips

• Leave it on by default. Smart Router is designed to balance quality and speed without any effort on your part, and we recommend setting it as your default model.
• Override whenever you want. If you want a specific model for a task, simply change the model from the model selector.
• Steer with effort. Set a higher effort level when you want deeper reasoning; the router will pick a model that supports it and score the request accordingly.
• See how it decided. Open the Message Info sidebar on any response from the Pentos Smart Router to see which model answered. Its Smart Router metadata view shows the full decision, including the difficulty score, the detected task, and the capability bar that was applied.
• See your energy savings. Pentos estimates the energy of every response. Open Energy Usage to see how much routing to a lighter model saves.
• Pair with a context mode. Combine Smart Router with Context Modes for the best balance of performance, cost, and energy.
• Not compatible with user API keys. At this time, BYOK is not supported with the Pentos Smart Router. If you want to use your own API keys, you’ll need to select the model manually and select your key. Learn more.