Compare the leading LLM routers, learned selectors, AI gateways, and cloud-native tools. This guide separates quality prediction from rule-based fallback so teams can evaluate token cost alongside completed work, retries, model coverage, and operational control.
The short version
Learned selectors predict which model can meet a quality target for each request. Rule-based gateways execute policies you define, such as budgets, metadata conditions, retries, and provider failover. Both can lower spend; the meaningful difference is whether the system can estimate model suitability before it routes.
At a glance
Weave Router
Learned per-action selection for engineering teams. Runs across coding harnesses and connects model choice to code quality, delivery, and cost per merged pull request.
OpenRouter Auto Router
A managed multi-provider endpoint with automatic selection options. It is a strong fit when catalogue breadth and fast provider switching matter more than a documented, task-specific quality signal. Verify the current Auto Router pool, strategy behavior, and session stickiness directly with OpenRouter.
Cursor Router
Cursor-native routing across the editor, agents, CLI, SDK, and mobile app. It offers Auto Cost, Auto Balance, and Auto Intelligence modes, but the policy remains inside Cursor rather than applying across separate coding harnesses such as Claude Code or Codex.
A lightweight, model-agnostic routing layer for teams that want routing separate from broader engineering analytics. The available product information does not document whether Ramp uses a learned quality model, fixed rules, or another selection method—so buyers should validate its decision logic, supported pool, and evidence model.
LiteLLM Router
A rule-based AI gateway rather than a learned quality selector. It balances traffic, retries failed calls, and supports configured fallbacks across 100+ providers through an SDK or OpenAI-compatible proxy. Best for platform teams prepared to define and operate the policy themselves.
RouteLLM
An open framework for training and evaluating a strong-versus-weak model selector. Its default preference-trained router estimates when a prompt merits the stronger model and exposes cost-quality trade-offs. Best for research-minded teams with the resources to validate, host, and maintain their own routing pipeline.
Not Diamond
A managed learned meta-router that predicts which candidate model should handle a request. It is useful when a team wants learned selection without building its own classifier. Trade-off: adding models may require retraining, and buyers should confirm how much intermediate decision evidence is exposed.
Portkey AI Gateway
A conditional gateway that applies sequential rules to request metadata, model parameters, and paths—then routes unmatched traffic to a default. It combines conditional routing with caching, guardrails, rate limits, budgets, key management, and 250+ models, but it does not predict answer quality.
Microsoft Foundry Model Router
An Azure-hosted learned router that selects a model per request in quality, cost, or balanced mode. Teams can constrain the candidate set and use automatic failover, including models from multiple vendors. Strong for Azure-first applications; less useful when cloud-neutral deployment is a hard requirement.
Amazon Bedrock Intelligent Prompt Routing
An AWS-native serverless router that predicts model performance and routes requests inside a supported model family, including Claude, Llama, or Amazon Nova options. It keeps invocation and billing inside Bedrock, but does not freely route across unrelated families and exposes limited public decision detail.
Multi-harness engineering teams
Choose Weave Router when you need one policy across Claude Code, Codex, opencode, and other supported endpoints—and want routing evaluated against code quality, pull-request keep rate, delivery outcomes, and cost per merged pull request.
Platform and infrastructure teams
Choose LiteLLM when you want a self-hosted, OpenAI-compatible proxy with broad provider support, provider-aware fallbacks, virtual keys, and operational control. Choose Portkey when conditional routing, guardrails, caching, spend controls, and enterprise gateway functions are central.
Azure or AWS-committed teams
Choose Microsoft Foundry Model Router when Azure is the operating platform and you want quality, cost, or balanced modes across a configurable Foundry model subset. Choose Bedrock Intelligent Prompt Routing when AWS-native serverless routing inside supported model families is the priority.
Router research and custom experimentation
Choose RouteLLM to own the research and evaluation loop for strong-versus-weak model selection. Choose Not Diamond when you prefer a managed learned meta-router, but confirm its current model coverage, benchmarks, and decision visibility with the vendor.
Give your teams the data they need to build the products you want.
Trusted by engineering teams from startups to Fortune 500
