OpenRouter

Overview

OpenRouter is an AI model aggregation API platform that provides a unified interface to access over 400 models from more than 70 providers. Developers only need one API Key to freely switch between different models—without having to register accounts with each model vendor separately.

Another value of OpenRouter is transparent price comparison and automatic routing—it displays real-time prices for each model, allowing you to choose the most cost-effective model as needed. Its catalog keeps expanding: on 2026/10/8 Mistral Large 4 entered public beta (1T total and 49B active parameters, native multimodality, 512K context, up to 256K output) alongside the GPT-6 Luna Decisions API, and on 2026/10/2 Black Forest Labs FLUX 3 Image arrived as a flagship image model for text-to-image and multi-reference editing rendering natively up to 4K. Beyond text, OpenRouter now covers embeddings (a 2026 guide covering 37 catalog entries, published 2026/9/23) and asynchronous video generation via POST /api/v1/videos with models such as Seedance, Veo and Wan, where switching models only requires changing the model identifier.

Key Features

• Unified API: Access over 400 models from more than 70 providers with one API Key, OpenAI SDK ready out of the box
• Transparent Pricing: Real-time display of API prices for each model, clear at a glance
• OpenAI Format Compatibility: API format compatible with OpenAI SDK, zero switching cost
• Model Routing: Can set automatic selection of the cheapest or fastest model, with failover to other providers; the new Auto router (openrouter/auto) launched 2026/8/10 matches models actually consumed by the community over the past 7 days across roughly 30 task types and supports the cost_tier parameter
• Free Models: Some open-source models offer free quotas
• Usage Monitoring: Detailed usage statistics and consumption breakdown; the Activity dashboard and beta Analytics API released 2026/8/17 show spending, token volume and cache hit rate by agent, model and request, with drill-down to individual request logs
• Custom Data Policy: Fine-grained control over data flow, ensuring prompts are only sent to trusted models and providers
• High Availability: Distributed infrastructure automatically falls back to other providers when one fails
• Batch API: Launched 2026/9/22, bills submitted batch inference requests at half the standard rate for offline processing and high-throughput jobs, using the existing OpenAI-compatible interface
• Video Generation API: Unified asynchronous video generation API: submit tasks via POST /api/v1/videos, poll status and download MP4 files; supports models such as Seedance, Veo and Wan, and switching models only requires changing the model identifier
• Image Models and Benchmarks: FLUX 3 Image (2026/10/2) handles text-to-image and multi-reference editing, rendering natively up to 4K; on 2026/9/18 OpenRouter benchmarked 20 image generation models with identical prompts, comparing billing, editing capability and output quality
• Embedding Model Guide: A 2026 guide published 2026/9/23 covers 37 embedding entries in the catalog with recommendations by task type and cost, aimed at retrieval, clustering and RAG workflows
• Evaluation Tooling: Ori Eval (2026/7/31) scans codebases, automatically writes eval files and runs evaluations to select the best model for a project from over 500 models; a 2026/9/14 tutorial covers using an LLM-as-a-judge to automatically score AI agent outputs against custom criteria
• Classifiers Beta: Launched 2026/7/21, automatically tags AI request usage and cost attribution

Use Cases

• Developers who need to flexibly switch between multiple AI models
• Comparing effects and cost-effectiveness of different models
• Model backend for AI applications (not locked to a vendor)
• Low-cost AI API for personal projects
• Model backend for local AI tools (e.g., Open WebUI)
• Offline processing and high-throughput batch inference jobs
• Retrieval, clustering and RAG workflows using embedding models
• Text-to-image and multi-reference image editing
• Asynchronous video generation pipelines
• Automatically evaluating AI agent outputs with an independent judge model

Pros

• One-stop multi-model: no need to manage multiple API Keys
• Transparent pricing: model prices clear at a glance
• OpenAI compatible: zero switching cost
• Free models: some models available at no cost
• Flexible switching: change models anytime without modifying code
• High availability: automatic failover ensures service stability
• Custom data policy: protect organizational data security
• Batch API at half the standard rate for high-throughput jobs
• Unified asynchronous video generation API alongside text, image and embedding models
• Activity dashboard and Analytics API for usage tracking by agent, model and request
• Ori Eval and LLM-as-a-judge tooling for model selection and output scoring

Pricing

OpenRouter does not charge a platform fee; billing is based on the actual API price of the model. Different models have different prices, and the platform regularly passes through vendor cuts on top of its own discounts: since 2026/7/30 GPT-5.6 Luna input is $0.1/M and output $0.6/M, while GPT-5.6 Terra input is $1/M and output $6/M. Mistral Large 4, live in public beta on 2026/10/8, is at half price for the first two weeks at $0.68 per million input, $2.09 per million output and $0.07 cached. A Batch API launched 2026/9/22 bills submitted batch inference requests at half the standard rate for offline processing and high-throughput jobs over the existing OpenAI-compatible interface. Some open-source models offer free quotas. Pay as you go, no minimum spend, no subscription required.

Summary

OpenRouter is the 'unified entry for AI APIs'—one account to use all models, transparent pricing, and on-demand switching. It is especially suitable for developers and AI enthusiasts who need flexibility in choosing between different models. If you only use one model, using the vendor's API directly may be more straightforward; but if you need flexibility, OpenRouter is the best solution. The platform now spans text, image, embedding and video models, and its routing layer has matured: the new Auto router (openrouter/auto) launched 2026/8/10 draws on weekly community consumption data of over 55T tokens, matches models actually consumed over the past 7 days across roughly 30 task types, supports the cost_tier parameter and respects account privacy settings. Evaluation tooling has grown alongside it, with Ori Eval (2026/7/31) scanning codebases, writing eval files and running evaluations to pick the best model for a project from over 500 models.

Version History

• Security Center launches (2026-10-23): OpenRouter launched a Security Center that audits every API key and grades it by risk, flags keys that can be removed, controls IP allowlists and adds spend alerts, periodic limits and maximum key lifetimes.
• Microsoft-Decision-1 decision model goes live (2026-10-09): Microsoft decision model Microsoft-Decision-1 went live on OpenRouter. Post-trained from Qwen3.5-9B, it is claimed to top about 36 blind benchmarks covering roughly 150,000 questions and to run 4.5x faster than the runner-up. Pricing is $0.042 per million input tokens with free output and a 32K context.
• Mistral Large 4 and GPT-6 Luna Decisions API go live (2026-10-08): OpenRouter brought up Mistral Large 4 in public beta and then the GPT-6 Luna Decisions API. Mistral Large 4 arrives with 1T total and 49B active parameters, native multimodality, a 512K context and up to 256K output, at half price for the first two weeks with $0.68 per million input, $2.09 per million output and $0.07 cached. The Decisions API lets an application choose between models, tools or actions and returns typed answers with probabilities
• FLUX 3 Image available (2026-10-02): Black Forest Labs FLUX 3 Image is now on OpenRouter as a flagship image model for text-to-image and multi-reference editing, rendering natively up to 4K
• OpenRouter publishes its 2026 embedding model guide covering 37 catalog entries (2026-09-23): OpenRouter published a 2026 guide to picking embedding models, covering 37 embedding entries in its catalog with recommendations by task type and cost, aimed at retrieval, clustering and RAG workflows.
• OpenRouter launches a Batch API at half price (2026-09-22): OpenRouter launched a Batch API that bills submitted batch inference requests at half the standard rate, aimed at offline processing and high-throughput jobs, using the existing OpenAI-compatible interface.
• OpenRouter benchmarks cost, editing and quality across 20 image generation models (2026-09-18): OpenRouter benchmarked 20 image generation models routed through its platform using identical prompts, comparing billing and output. Cost per image ranged from 0.006 to 0.134 dollars, a spread of roughly 22 times, with differences in editing capability and image quality also covered.
• OpenRouter Tutorial: Using LLM-as-a-judge to Automatically Evaluate AI Agent Outputs (2026-09-14): OpenRouter has released a tutorial explaining how to use an independent judge model to automatically score AI agent outputs according to custom scoring criteria, covering the quality of open-ended responses that deterministic tests cannot check.
• OpenRouter video generation API: a code-first integration guide (2026-08-25): OpenRouter launches a unified asynchronous video generation API, where tasks are submitted via POST /api/v1/videos, status is polled, and MP4 files are downloaded. It supports models such as Seedance, Veo, and Wan, and switching models only requires changing the model identifier.
• OpenRouter announces it is joining Stripe (2026-08-19): OpenRouter announces a merger with Stripe to accelerate global economic growth. OpenRouter currently processes over 10+ trillion tokens daily from 400+ AI models, serving more than 10 million developers and companies, with inference volume growing at least 10x annually since its founding. After the merger, OpenRouter will continue to operate independently under its original name and mission, with no changes to its product or roadmap, and routing decisions will remain prioritized by user interests. The transaction is expected to close within the coming weeks.
• OpenRouter launches Activity dashboard and Analytics API: track AI usage by agent, model and request (2026-08-17): OpenRouter releases Activity dashboard and beta Analytics API, allowing users to view metrics such as spending, token volume, and cache hit rate by agent, model, and request dimension, with support for drilling down to individual request logs.
• OpenRouter launches a real-time web search benchmark: how to choose engine, depth and model for your agent (2026-08-12): OpenRouter released a real-time leaderboard that systematically evaluates four configuration combinations: model, search engine, search method, and budget. Data shows that increasing the search budget from 1 round to 25 rounds nearly doubles the BrowseComp score, with costs rising only 2.5-7 times; model selection matters more than the engine, with an average score difference of 15 points vs. 10 points. For tasks with high failure rates, search depth should be reduced to control costs.
• OpenRouter launches a new Auto router powered by marketplace intelligence (2026-08-10): Based on weekly community consumption data of over 55T tokens, OpenRouter has launched a new version of the Auto router (openrouter/auto), whose model selection outperforms the old version across most tasks and cost tiers. The new router matches models actually consumed by the community over the past 7 days across approximately 30 task types, supports the cost_tier parameter (from low to max), and adheres to account privacy settings. On benchmarks such as MMLU Pro, the new default tier achieves performance equivalent to the old version in most domains at a lower cost.
• OpenRouter launches Ori Eval to simplify evaluation (2026-08-03): Introducing Ori Eval: the simplest way to write your first eval. There is no single best model, only the model that fits each task best. Ori Eval uses OpenRouter's API to run every task in your codebase, then evaluates the results. curl -fsSL https://openrouter.ai/skills/spawn-ori-eval
• OpenRouter launches Ori Eval: prove with your own data which model fits your project best (2026-07-31): OpenRouter releases Ori Eval, an agent that can scan codebases, automatically write eval files, and run evaluations, helping developers select the single best model for their projects from over 500 models.
• OpenRouter cuts prices for GPT-5.6 Terra/Luna (2026-07-30): GPT-5.6 Terra and Luna just got a price cut from @OpenAI. OpenRouter's prices are still lower. Our 50% exclusive discount applies on top of that, bringing Luna input down to $0.1/M and output to $0.6/M, and Terra input down to $1/M and output to $6/M. Scale your workloads with luna: https://openrouter.ai/openai/gpt-5.6-luna
• Launched Gemini 3.6 Flash + Classifiers (2026-07-21): Launched Gemini 3.6 Flash and 3.5 Flash-Lite (150+ tok/s); Classifiers beta automatically tag AI request usage and cost attribution