Weave Router 2.0

Product Introduction

1. Definition: Weave Router 2.0 is an intelligent, open-source model routing layer specifically designed for AI-powered coding agents and developer tools. It functions as a dynamic cost-optimization proxy that sits between your application and multiple large language model (LLM) providers.
2. Core Value Proposition: It exists to dramatically reduce LLM inference costs—by 30-60%—for software development workflows while maintaining or exceeding the quality output of frontier models like GPT-6 Astra. Its primary value is delivering "Astra-level quality at half the cost" through intelligent, per-request routing.

Main Features

1. Intelligent Per-Turn Model Routing: The system analyzes every individual request (or "turn") from a coding agent in real-time. It uses a proprietary complexity classifier, trained on extensive session data, to score the task's difficulty in milliseconds. Based on this score and real-time pricing, it routes the request to the cheapest capable model in its configured pool (e.g., DeepSeek, Llama, GPT, Claude).
2. Cache-Aware Switching Logic: Unlike naive routers, Weave Router 2.0 incorporates an "ache-aware" switching mechanism. It calculates whether the cost savings of switching to a cheaper model outweigh the computational cost of rebuilding the LLM's context cache. This prevents wasteful switching that could increase latency or cost.
3. Live Performance Monitoring and Escalation: An integrated escalation classifier continuously monitors the performance of the currently engaged, cheaper model. If it detects signs of stalling, looping, or failing to complete the task correctly, the system automatically and seamlessly escalates the request to a more capable frontier model to ensure task completion.
4. Subscription and Quota-Aware Load Balancing: The router intelligently manages API keys and subscription quotas. It can utilize models from different subscriptions (e.g., using Claude models within a Codex plan or GPT models within a Claude Code plan), draining flat-rate subscription quotas first before consuming pay-as-you-go API credits, maximizing existing investments.

Problems Solved

1. Pain Point: Prohibitive and unpredictable costs of using state-of-the-art (frontier) LLMs like GPT-4, Claude 3.5 Sonnet, or GPT-6 Astra for all coding agent tasks, many of which are simple and do not require top-tier model capability.
2. Pain Point: Inefficient manual model selection where developers must constantly choose between cost and performance for different coding tasks (e.g., refactoring vs. debugging vs. documentation).
3. Target Audience: Engineering teams and developers using AI coding assistants (Claude Code, Codex, Cursor, etc.); DevOps and platform engineers managing AI tooling budgets; startups and enterprises scaling AI-powered development workflows.
4. Use Cases: Automatically routing simple code summarization or YAML migration to cost-effective models like Llama; reserving expensive models like Claude 3.7 for complex tasks like debugging flaky integration tests; optimizing spend across team-wide deployments of AI coding assistants.

Unique Advantages

1. Differentiation vs. Competitors (e.g., OpenRouter Auto): Weave Router 2.0 demonstrates superior quality-cost trade-offs. On benchmarks like Terminal-Bench 4.0, it matched GPT-6 Astra's quality at 48% lower cost, whereas OpenRouter Auto solved far fewer tasks (25% vs 62.1%). Its "ache-aware" switching and live escalation provide a more robust and reliable routing logic compared to simpler cost-based routers.
2. Key Innovation: The dual-classifier system (complexity + escalation) combined with cache-aware economics. This multi-question approach—assessing difficulty, switch cost, and live performance—creates a dynamic routing policy that optimizes for total cost of completion, not just per-token price, which is a significant technical advancement in LLM orchestration.

Frequently Asked Questions (FAQ)

1. How does Weave Router 2.0 save money on LLM API costs? It saves money by dynamically routing each coding task to the least expensive large language model capable of correctly completing it, using a real-time complexity classifier to avoid over-provisioning expensive frontier models for simple requests.
2. What is the performance impact of using a model router like Weave? According to published benchmarks on SWE-Atlas, Weave Router 2.0 can be 2.5x faster per trial than GPT-6 Astra while costing 54% less, as faster, cheaper models handle simpler tasks, reducing overall latency and cost.
3. Is Weave Router 2.0 compatible with Claude Code and Cursor? Yes, Weave Router offers one-command integration that automatically detects and configures itself for popular AI coding environments including Claude Code, Codex (formerly Windsurf), and Cursor, routing their requests through its optimization layer.
4. How does the "ache-aware switching" in Weave Router work? Ache-aware switching is a cost-benefit algorithm that prevents the router from switching to a cheaper model if the computational cost of rebuilding the new model's context cache would be greater than the potential token savings, ensuring net-positive cost optimization.
5. Can I self-host Weave Router 2.0? Yes, Weave Router 2.0 is open-source and released under the Elastic License 2.0 (ELv2), allowing organizations to self-host the entire routing infrastructure for complete control, data privacy, and further customization.