Automatic routing
One box. No model picker.
Choosing between AI models is work you should not have to do. Ask Discrezo once and the router decides which approved model is best suited to the question, then gets out of the way.
Different question. Different model. Same assistant.
The decision, live
You ask once. The router does the rest.
You ask
“Review this contract and tell me which clauses I should negotiate.”
Discrezo router
Understanding the task...
Task
Analysis
0Difficulty
High
1Sensitivity
Standard
2Context
Long document
3Routed
Waiting
The answer streams as soon as the selected model begins responding.
The route is chosen per question, not per account.
Your question
Router
Open-weight models
Everyday work
Routine and straightforward tasks.
Frontier models
When the task warrants it
Harder tasks where frontier capability matters.
The route is chosen per question, not per account.
Two modes
Standard by default. Private when you want it.
Standard
Best model for the job.
Standard Mode allows Discrezo to route across approved open and frontier models according to the task.
- Harder code, analysis, writing and long-context work can be routed to models with the capabilities needed for the job.
- AI providers receive the prompt and context needed to answer.
- Discrezo does not attach your account ID, email, payment identity or originating network identity to the request.
- Routes are limited to providers that meet Discrezo's approved privacy requirements.
- Straightforward questions do not automatically need the most expensive or powerful frontier model.
Default behaviour: optimise for answer quality, speed and suitability.
Private mode
Open models only.
Turn on Private Mode and routing is constrained to approved open-weight inference.
- Prompts do not go to OpenAI, Anthropic or Google.
- The privacy constraint is applied before model quality is considered.
- There is no fallback to a major proprietary provider while Private Mode is active.
- The interface, history and automatic routing experience otherwise remain the same.
Private Mode may give up some capability on the hardest tasks in exchange for stricter provider choice. That trade-off is yours to make.
The decision
Four signals. Then a route.
The router evaluates the request itself. The inference side does not need your Discrezo account identity to make the decision.
01
Task
What kind of work is this?
Chat, writing, code, analysis, current-information requests. Different types of work favour different model capabilities.
02
Difficulty
How much capability does the question really need?
A simple rewrite does not need to take the same route as a difficult coding or reasoning task. Frontier capability is used where it materially helps.
03
Sensitivity
Does this request need a more restrictive route?
Privacy constraints are applied before ordinary quality ranking. Private Mode is an explicit hard constraint, and sensitive requests can also narrow the eligible route set.
04
Context
How much information must the model handle?
Long documents and large conversations narrow the set of models that can answer reliably.
The router evaluates the question. It does not need to know whose account asked it.
Identity and inference remain separate.
Different work. Different routes.
One assistant can make different decisions.
Rewrite this email
Debug this production error
Analyse this 80-page contract
Summarise this meeting
GPT · Claude · Gemini · leading open models
Design principle 01
Model names are our problem.
Most people do not want to become experts in model benchmarks, context windows and provider release cycles. They want a good answer.
Discrezo has one text box in the default experience. No model marketplace. No model picker. No need to learn which provider released the best coding model this month.
You ask. We route.
Design principle 02
Routing should feel invisible.
The router is designed to make its decision while the rest of the request is already being prepared. It should not feel like another step between asking and answering.
Answers stream as soon as the selected model begins responding.
Routing overhead target
under 100 ms at p95
Engineering target, not a consumer latency guarantee.
Design principle 03
If a provider has a bad day, you shouldn't.
The router considers provider availability as well as model suitability. If a chosen route is unhealthy, Discrezo can move to the next approved route without asking you to start again.
You are not expected to manage provider outages.
Design principle 04
Predictable before clever.
The launch router starts from explicit, testable routing rules. More complex learned routing only earns its place if real product data shows that it improves outcomes.
We would rather understand why a request took a route than add machine learning for the sake of saying the router uses machine learning.
The “best” model is not a leaderboard.
Quality
Can it answer the task well?
Capability
Can it handle the required context and tools?
Speed
Can it respond quickly enough for the job?
Availability
Is the route healthy right now?
Private Mode adds another hard constraint: provider eligibility.
The seam
The router needs the question. It does not need your identity.
The inference path receives the prompt and relevant context needed to make a routing decision and generate an answer. It does not receive the Discrezo account identity belonging to the user who asked it.
Account side
Who has access
Subscription entitlement
Does not read the prompt.
Routing side
The task
The context
The eligible models
Does not receive the Discrezo account identity.
Discrezo knows who, not what. The routing side knows what, not who.
The test that matters
Routing is only useful if the answers hold up.
A router that saves money but gives worse answers is not a better product. Neither is a router that adds noticeable delay just to choose between models.
Our standard is simple: automatic routing should hold its own against a strong single-model baseline on quality while improving the experience on everyday work through speed, suitability and automatic provider choice.
We test the router against fixed evaluation sets before treating routing changes as improvements.
What stays internal
The router is proprietary. We explain what it considers and what guarantees it must respect. We do not publish: internal scoring, routing thresholds, provider ranking, route tables, failover graphs, classifier implementation, provider economics, evaluation datasets.
Transparency about behaviour does not require publishing the recipe.
One box
Ask the question. Let Discrezo choose the AI.
One assistant across leading AI models. Automatic routing by default. Private routing whenever you want it.
Founding members lock $8/month for life. No card required.