Skip to main content
Forge Router selects the model and reasoning effort for each request across your configured model providers. It is designed for agent workloads where different steps in the same workflow can require different levels of speed, capability, and depth. Router decisions are optimized for the completed task, not only for the cost or latency of an individual model request.

​
Availability

Forge Router is currently available in beta through Forge LLM Gateway. Add the managed Forge Router destination to an access profile, then send requests to that profile using the model name forge/auto. Forge selects from the model portfolio reviewed for your organization; applications do not need to choose a specific upstream model for each request. Forge Router must be the only destination in its route tier. Use later priority tiers when you want a separate fallback path.

​
What Router considers

Router evaluates the request and its surrounding session context against the available model portfolio. The decision can account for: The result is not a fixed ranking of models. A model that is effective for one task category may be inefficient or insufficient for another.

​
Company-specific optimization

Router recommendations improve as Forge observes more of your organization’s work. Signals associated with completed tasks can include durable outputs, verification, review outcomes, retries, manual rework, and other available evidence of task quality. This allows Router to learn patterns specific to your company, such as:
  • Which models perform reliably for recurring task types.
  • Which tasks need additional reasoning effort.
  • Which model choices reduce retries or downstream correction.
  • Which workflows benefit from keeping a consistent model path.
  • Where a lower-cost or faster model produces comparable results.
Recommendations are based on observed task outcomes rather than only prompt length, keywords, or generic benchmark results. As models, pricing, and usage patterns change, the routing policy can adapt to the new evidence.

​
Agent-session awareness

Agent sessions are sequences of related requests, tool calls, and intermediate outputs. Router considers the broader workflow so that an optimization for one request does not create unnecessary cost or latency elsewhere in the session. This includes prompt-cache awareness. Long sessions may reuse a large context across requests, and changing providers or model paths can reduce the value of that cached context. Router weighs the benefit of selecting a different model against the benefit of maintaining cache locality and session continuity.

​
Routing outcomes

Router is intended to improve the balance of:
  • Output quality and task completion
  • Reasoning depth and response latency
  • Inference spend and session efficiency
  • Model flexibility and workflow consistency
The effectiveness of a routing decision can be evaluated through the same signals used by Forge Optimize, including verification, rework, durable artifacts, latency, and cost. This provides a basis for comparing routing behavior against the actual outcomes of your AI workloads.

​
Model portfolio

Router works across the models and providers available to your organization. Adding or removing a model changes the set of eligible routes; observed task outcomes continue to inform which available model is best suited to each class of work.