OpenRouter
Free
OpenRouter is the industry's largest unified LLM API gateway, aggregating 400+ models and 70+ inference providers in the AI DevOps domain. Through mechanisms such as smart price routing, automatic failover prompt caching, and Zero Data Retention, it provides developers with an OpenAI-compatible single point of access, eliminating vendor lock-in, reducing inference costs, and improving availability.
OpenRouter: In-depth teardown of unified LLM API gateway
OpenRouter is not a standalone large model, nor is it a chat application - it is an "intelligent switch" connecting AI applications to 400+ large models. Since its founding in early 2023, it has evolved from the earliest LLM market to today's largest unified AI gateway in the world, processing 100 trillion tokens monthly, serving more than 10 million global developers, and receiving joint investment from top institutions such as CapitalG (Alphabet), a16z, Menlo Ventures and Sequoia Capital.
Tool type determination: Main type - [Basic large model / API infrastructure] (unified inference API gateway, aggregated multi-provider model access); secondary type - [Agent / MCP / Automation tool] (provides Agent SDK and MCP Server, supports tool invocation and automated workflow).
Core parameters and statistics
| Project | Current Public Information |
|---|---|
| Product positioning | Unified LLM API gateway, aggregating multi-provider reasoning capabilities |
| Delivery form | API endpoint (/api/v1/chat/completions) + Web console + Chat interface |
| Model scale | 400+ active models (covering text, image, audio, video, embedding, reordering and other modalities) |
| Number of providers | 70+ inference providers (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, etc.) |
| Monthly processing volume | 100 trillion tokens |
| Global users | 10 million+ |
| Access applications | 250,000+ applications, covering 4.2 million+ global users |
| API compatible | Fully compatible with OpenAI Chat Completions format |
| Official SDK | TypeScript (@openrouter/sdk), Python (openrouter), Agent SDK (@openrouter/agent) |
| MCP support | Official remote MCP Server: https://mcp.openrouter.ai/mcp |
| Company establishment | Early 2023 |
| Headquarters/Region | US |
| Investors | CapitalG (Alphabet), a16z, Menlo Ventures, Sequoia Capital |
A brief comment: OpenRouter encapsulates all four problems that should be solved by the infrastructure layer: "selecting models, changing providers, managing costs, and ensuring availability" with an OpenAI-compatible API endpoint.
Interpretation of key parameters: 400+ models are not statically displayed, but are dynamically available through real-time uptime monitoring and price sorting. 100 trillion monthly tokens means that OpenRouter is already one of the key hubs for global AI inference traffic, and its routing algorithm is continuously iteratively optimized in real traffic.
User and market recognition
Public data signals
- Official Disclosure: Monthly processing of 100T tokens, 10 million+ global users, 250,000+ application access - these data come from the about page and homepage of the official website, and are publicly verifiable first-level information.
- Financing endorsement: In May 2026, the New York Times reported that OpenRouter completed US$113 million in financing, and investors included CapitalG (an investment institution of Alphabet); during the same period, the Wall Street Journal reported that it completed US$40 million in financing. The investment lineup includes CapitalG, a16z, Menlo Ventures, Sequoia Capital - this is the top allocation in the large model infrastructure track.
- Developer Community: OpenRouter has active communities on GitHub (OpenRouterTeam), Discord, X/Twitter and other channels, and officially provides a Status Page (status.openrouter.ai) to monitor service status in real time.
Market benchmarking analysis
| Comparison dimensions | OpenRouter | Directly call a single provider API | Build your own multi-model routing layer |
|---|---|---|---|
| Number of connected models | 400+ (out of the box) | 1 | Depends on engineering investment |
| Failover | Automatic (based on real-time uptime data) | None | Self-research required |
| Price optimization | Default price load balancing + optional manual sorting | Fixed price | Self-built bidding logic required |
| Maintenance cost | Zero (hosted gateway) | Low | High (requires dedicated team) |
| Prompt cache | Multi-provider automatic/explicit cache support | Provider dependent | Need to build your own |
| Data policy control | Grayscale control ZDR mandatory, provider tag | Depends on provider | Need to build by yourself |
Boundary Statement: The above financing information comes from NYT/WSJ public reports, and the specific valuation and terms have not been disclosed. If enterprise procurement requires SLA terms, compliance certification (SOC2, etc.) or enterprise customer cases, you need to directly contact the OpenRouter business team to obtain the latest materials.
Cost advantage
C-side/individual developer
- Free Tier: 25+ free models (like
nvidia/nemotron-3-embed-1b:free), 4 free providers, 50 daily request limit. Suitable for prototype verification and personal learning. - Pay-as-you-go: No minimum spend, no lock-in. The platform price increase is only 5.5% (based on the provider's list price), which is far lower than the engineering and operation and maintenance costs of self-built multi-provider access.
- BYOK (bring your own key): The first $25,000 of list price inference volume per month is free of platform fees, and only 5% is charged for the remainder.
Developer/API calls
- Hidden Cost Savings: Automatically select the lowest price provider through intelligent routing, and often get better prices than a single provider with the same model capabilities. Popular models such as DeepSeek V3 are deployed across multiple providers, and OpenRouter defaults to price-weighted routing.
- Caching Savings: OpenRouter's Prompt Caching saves 75%-90% on input tokens at providers that support cached reads (Anthropic 0.1x, DeepSeek 0.1x, Google 0.25x, etc.).
- NO MINIMUM: Prepaid credits are available across all models and providers, with no monthly minimum commitment.
Enterprise level
- Platform Rate Discount: Enterprise contracts can negotiate a lower platform markup ratio.
- BYOK Extension: Free platform fee for the first $200,000 of list price inference volume per month, 5% for the remainder.
- Optional Dedicated Rate Limiting: Avoid multi-tenant contention for resources.
- Invoice Payment: Supports corporate settlement.
Free Truth: The free tier’s 50-shot/day limit is insufficient for production-level applications; free models are typically smaller or quantized versions that are not suitable for complex inference. Although there is no minimum consumption for Pay-as-you-go, the 5.5% platform fee is a considerable expense under extremely large-scale calls, and needs to be compared with the discounted price of the enterprise's direct purchase provider.
Hidden costs: Integrating OpenRouter itself costs almost nothing (OpenAI compatible SDK plug-and-play), but requires the team to understand its routing semantics (provider, order, sort, data_collection and other parameters) to realize the full value. Configurations such as enterprise-level SSO, audit logs, and data policy governance also require additional initial configuration investment.
Main functions
1. Unified AI inference API
Through an OpenAI compatible /api/v1/chat/completions endpoint, you can access 400+ models covering text generation, image understanding and generation, audio/speech, video understanding, embedding, reordering, code completion and other modalities. Developers do not need to register accounts with each provider, manage multiple API keys and different SDKs.
2. Intelligent provider routing
OpenRouter’s core differentiating capabilities. Defaults to load balancing by price (inverse square of price), taking into account provider availability in the last 30 seconds. Supports multiple custom routing strategies:
- Manual ordering (
provider.order): Specify provider priority list - Performance sort (
provider.sort): Sort byprice,throughput,latency - Performance Thresholds (
preferred_min_throughput/preferred_max_latency): Set percentile throughput/latency preference - Partition routing (
sort.partition):model(default, group by model) ornone(sort globally across models) - Shortcut:
model:nitro(sorted by throughput),model:floor(sorted by price)
3. Automatic failover and uptime optimization
Track response times, error rates, and availability across all provider endpoints in real time. When the primary provider fails, it automatically switches to the sub-optimal provider, leaving the user unaware. Historical uptime charts are publicly available, and the service status page is updated in real time.
4. Prompt Caching (multi-provider unified abstraction)
Unified abstraction of caching mechanisms from providers such as Anthropic (explicit cache_control, 5 minutes/1 hour TTL), OpenAI (automatic + explicit cache), DeepSeek (automatic), Google Gemini (implicit + explicit cache), Z.AI (automatic), Grok (automatic), Moonshot AI (automatic), Groq (automatic), Alibaba Qwen (explicit cache), etc., developers only need to add in the request cache_control tag, OpenRouter automatically maps to each provider's cache semantics. Automatically enable Provider Sticky Routing to keep the same session routed to the same provider by session_id to maximize cache hit ratio.
5. Data policy and security control
- Zero Data Retention (ZDR): Force routing at request granularity to endpoints that do not retain hint data
- Data Collection Policy: Exclude providers that store user data via
data_collection: 'deny' - Distillable Text Enforcement:
enforce_distillable_text: trueonly routes to models that allow text distillation - Provider-Specific Headers: Transparently transmit Anthropic Beta Features (such as Interleaved Thinking, Structured Outputs)
6. Agent SDK and MCP Server
- Agent SDK (
@openrouter/agent): Define tools based on Zod schema, support multi-round dialogue cycle, tool execution, status management, and complete the complete agent cycle of "send→tool call→result return→final response" through a single call ofcallModel. - MCP Server: Official remote MCP server (
https://mcp.openrouter.ai/mcp), zero installation configuration, allowing AI coding assistants (Claude Code, Cursor, Codex, etc.) to pull OpenRouter’s model list, price, balance, usage ranking and document search in real time.
7. Chat, Rankings, Labs and other web tools
- Chat: Online conversation interface, you can directly experience different models
- Rankings: Model usage ranking (sorted by weekly trend of tokens consumption)
- Models: Complete model browsing and comparison page
- Apps: Display of selected applications integrating OpenRouter
- Labs: Experimental feature preview area
Expert view: The real synergy of OpenRouter lies in the linkage of routing + caching + failover + data strategy**. Cache relies on Sticky Routing to maintain session affinity with the same provider; failover automatically switches to a suboptimal provider when the cache becomes invalid; data policy filters non-compliant providers before routing decisions. Developers don't need to worry about "whether this provider supports caching" and "whether that provider will store data" in the application code - these are all done once in the routing layer.
Model and version evolution
Current main line
OpenRouter continues to iterate in a SaaS model and does not have a public version number in the traditional sense. The evolution of its capabilities is reflected in model catalog expansion, routing algorithm optimization, security and data policy function enhancements, and SDK/toolchain updates.
Key Milestones
| Time | Milestone | Description |
|---|---|---|
| Early 2023 | Founded | Launched as the earliest LLM marketplace |
| 2024 | Scale expansion | The number of models continues to grow, and a price load balancing default routing strategy is established |
| Early 2025 | Early online service | Complete core gateway infrastructure and aggregate mainstream model providers |
| Early 2026 (v2 stage) | Auto Router, Prompt Caching, BYOK | Introducing intelligent routing, cache abstraction, and bring-your-own-key functions |
| Mid-2026 | MCP Server, Agent SDK | Expand to AI agent and development tool ecosystem |
| May 2026 | US$113 million in financing | CapitalG led the investment, significantly increasing valuation and strategic resources |
| July 2026 | 400+ models 70+ providers | Covering all modes including text, image, audio, video, embedding, etc. |
Version evaluation suggestions
For the production team, the focus should be on whether OpenRouter continues to support the latest and strongest required models, whether the routing strategy meets latency/cost requirements, and whether the data strategy is aligned with enterprise compliance requirements, rather than pursuing a fixed "version number". OpenRouter's Status Page and Changelog are the best channels to track service evolution.
Technical advantages
Architecture link
User Application → OpenRouter API (/api/v1/chat/completions)
↓
Routing engine (real-time uptime + price + performance data)
↓
┌─────────┼─────────┐
↓ ↓ ↓
Provider A B ... 70+
↓ ↓ ↓
Model A1 B1...Control flow: The user request carries model (or models array as fallback) + routing parameters (provider object). The OpenRouter routing engine integrates the provider's real-time health, price, performance threshold and user policy preferences, selects the optimal provider endpoint and initiates an inference call.
Data reflow: inference results + usage information (including cache hit details) → return to the user via the original route, while retaining traceable records on the Activity page.
Why OpenRouter is faster/more economical/more stable
- More economical: Default price load balancing, weighted routing according to the inverse square of price, market mechanism drives price competition. Multi-provider bidding drives popular model prices toward the lowest available levels.
- More stable: Real-time uptime tracking of 70+ providers, faults within a 30-second window are automatically included in routing decisions. Single provider failure does not affect overall service availability.
- Faster: Edge deployment (edge node close to user), optional
throughputsorting routing to the provider endpoint with the highest throughput.:nitroshortcut is enabled with one click. - More flexible: from "fully automatic" (default load balancing) to "full control" (manual
order+allow_fallbacks: false), supporting multiple routing dimension combinations from price, delay, throughput to data policy.
Guide to engineering pitfalls (a must-read for developers)
- Interaction of route parameters and fallback:
provider.orderspecifies the priority order, but fallback is enabled by default. If you want to "use only the providers specified in the list and no others", you must also setallow_fallbacks: false. Otherwise, OpenRouter will automatically fallback to the full provider pool when no provider in the list is available - this may be a security risk in some compliance scenarios. - Caching semantic differences: The caching mechanisms of different providers vary greatly. Anthropic explicitly
cache_controlneeds to set a breakpoint, OpenAI automatic caching takes effect automatically above 1024 tokens, Google Gemini needs to insertcache_controlbut only uses the last breakpoint. OpenRouter does the abstraction but there are still behavioral differences (e.g. Anthropic’s 5 minute vs 1 hour TTL cost is different). It is recommended to check the actualcached_tokensandcache_discountvalues when switching models or providers. - BYOK fee structure: Under BYOK mode, the first $25K (PAYG)/$200K (Enterprise) of list price reasoning per month is free of platform fees, and 5% is charged for the excess. Note that this is calculated based on "list price" rather than "actual consumption". If you get a discounted price from the provider through BYOK, the platform fee will still be calculated in proportion to the original list price, which needs to be calculated carefully.
- Model slug naming rules: OpenRouter uses the
provider/model-nameformat (such asanthropic/claude-sonnet-4.5).:nitroand:floorare suffix shortcut keys.~openai/gpt-latestis the latest alias (latest alias), which is automatically resolved to the latest flagship model. It is recommended that the production environment pin the specific version number (snapshot) to prevent the latest alias from accidentally switching to destructive changes.
How to use
Relevant information has not been made public, please refer to the official real-time page.
Product Pricing
Charging model
OpenRouter adopts the **pre-recharge + pay-as-you-go model and does not charge a monthly subscription fee. Users purchase Credits first (minimum $5), and the actual token consumption is deducted, and the limit is common across all models and providers.
Comparison of three-tier solutions
| Dimensions | Free | Pay-as-you-Go | Enterprise |
|---|---|---|---|
| Platform rate | None | 5.5% (based on list price) | Negotiable discount |
| Available Models | 25+ Free Models | 400+ All Models | 400+ All Models |
| Available Providers | 4 Free Providers | 70+ All Providers | 70+ All Providers |
| Rate Limit | 50 requests/day | High Global Limit | Optional Private Limit |
| BYOK free platform fee limit | — | $25K monthly list price | $200K monthly list price |
| Excess BYOK Rate | — | 5% | 5% |
| Payment methods | — | Credit card, cryptocurrency, etc. | Corporate invoice |
| Prompt Caching | Depends on model | ✓ | ✓ |
| Automatic routing/provider sorting | ✓ | ✓ | ✓ |
| Budget and Spending Control | — | ✓ | ✓ |
| Activity Log and Export | — | ✓ | ✓ |
| Manage API Key | — | ✓ | ✓ |
| SSO/SAML | — | — | ✓ |
| Contract SLA | — | — | ✓ |
| Support Channels | Community Support | Email Support | SLA + Dedicated Slack Channel |
Hidden costs and risks
- Platform markup: 5.5% may not seem high, but it is a significant expense under large-scale calls (monthly consumption $100,000+). At this time, the enterprise direct provider discount + Enterprise plan may be better.
- Model Price Fluctuation: OpenRouter's model price is set by the provider, and OpenRouter will adjust accordingly. Frequent price changes can affect cost forecasts.
- Availability of Free Models: Free models may be withdrawn or downgraded by the provider at any time and are not suitable for production context dependence.
Application scenarios
Scenario 1: AI application development and multi-model experiment
When developers build AI products, they need to compare the performance between multiple models (such as Claude vs GPT-5 vs Gemini). The traditional method requires maintaining multiple provider accounts and SDKs. OpenRouter reduces the switching cost to "changing a parameter" through a unified API. Suitable for: AI entrepreneurial teams, independent developers, product prototype verification.
Efficiency deduction: From "registering 5 provider accounts, managing 5 SDK versions, and writing 5 sets of fallback logic" to "1 API Key, 1 SDK, and 1 set of routing parameters". Engineering access time is reduced from 2-3 days to 1-2 hours.
Scenario 2: Production-level inference high-availability architecture
For real-time, user-facing AI products, single-provider failure = business outage. OpenRouter's automatic failover is based on real-time uptime data, detecting and switching within 30 seconds. Enterprises can control the failover scope through provider.order + allow_fallbacks, and achieve observability with Status Page and Activity log. Suitable for: Customer service robots, real-time translation AI search, code completion and other businesses that require high-availability reasoning.
Scenario 3: Cross-model cost optimization
Enterprises that operate large-scale AI inference (such as social media content generation, mass production of e-commerce copywriting) are automatically routed to the most economical provider endpoint through OpenRouter's price load balancing. Combined with Prompt Caching to reduce the cost of repeatedly entering tokens. Developers can also set max_price to prevent exceeding the budget. Suitable for: Scenarios of high-frequency AI calls in content platforms and e-commerce operation SaaS products.
Cost deduction (for reference only, unofficial commitment): For large vocabulary multi-turn dialogue scenarios, after enabling Prompt Caching + price routing, the effective token cost may be reduced by 40%-70% (depending on the cache hit rate and provider price distribution). Actual savings vary based on usage pattern.
Scenario 4: AI agent and automated workflow
Use the callModel of the Agent SDK to build an AI agent with tool calls to automatically complete multi-step reasoning tasks. Combined with MCP Server, the AI coding assistant can perceive available models and prices in real time to achieve smarter model selection. Suitable for: Automated operation and maintenance, code review agent, intelligent customer service workflow, automated testing.
Scenario 5: Data compliance-sensitive industry applications
Industries that are sensitive to data retention, such as finance, medical care, and legal affairs, can use data_collection: 'deny', zdr: true, and provider.only to limit requests to be routed only to providers that comply with data policies. SSO/SAML and audit logs for Enterprise plans further meet compliance requirements.
Not suitable for the scene
- Fully Offline Deployment: OpenRouter is a managed cloud service and does not support private deployment or air-gapped deployments.
- Custom model fine-tuning required: OpenRouter does not provide model training/fine-tuning capabilities.
- Extreme low latency scenario (< 50ms): The additional network hops introduced by the routing layer may increase latency. Although edge deployment alleviates this problem, it is not comparable to direct calls within the same data center.
- Ultra-large-scale calls (monthly consumption $1 million+): At this scale, the total platform fee of 5.5% is considerable. It needs to be compared in detail with the total cost of ownership of direct signing from the provider + self-built lightweight routing layer.
Applicable people
- Individual Developers & Independent Hackers: Validate ideas with the free tier, flexibly scale with PAYG, access 400+ models with one API Key, and no need to manage multiple provider accounts.
- AI Entrepreneurship Team: Rapid multi-model experiments to reduce engineering investment. Routing and failover capabilities enable small teams to deliver enterprise-class inference availability.
- Medium SaaS & Product Teams: Continuously optimize inference costs with price routing and Prompt Caching. Activity logs and budget controls help manage team spending.
- Enterprise Digital and AI Infrastructure Team: The Enterprise solution's SSO, audit logs, data policy routing, and contract SLA meet organizational-level governance needs.
- AI Agent/Agent Developer: Agent SDK and MCP Server provide a complete link from model invocation to tool execution.
Dissuade the crowd:
- Government and enterprise customers who require completely offline privatized deployment
- AI research teams that need deep customization of underlying model training/fine-tuning
- Enterprises with extremely strict data sovereignty requirements (must have domestic data centers) and OpenRouter does not cover this area
- Customers using a single provider with very high usage (direct contract may be cheaper)
Summary and Outlook
OpenRouter's core competitiveness lies in: Using an OpenAI-compatible API endpoint to encapsulate a full set of inference infrastructure capabilities from model selection, provider failover, price optimization to data policy governance. It's not "just another API provider", but a provider of API providers - providing developers with greater bargaining power, higher availability and more flexible routing control downstream by aggregating 70+ upstream inference services.
Current limitations: As a hosting service, there is an inherent delay overhead of network jumps; the platform fee ratio under ultra-large-scale calls needs to be optimized through enterprise negotiation; public information on compliance certification is limited.
Procurement/Adoption Risk Assessment:
- Low risk of vendor lock-in - OpenRouter is compatible with the OpenAI format, and switching back to the direct connection provider only requires changing
baseURL. - Medium price risk - The platform’s price increase of 5.5% is within control, but large-scale use requires regular comparison of the total cost of the direct signing plan.
- Low availability risk - A multi-provider redundant architecture is naturally better than a single point of dependence, but it should be noted that improperly configured routing policies may reduce the scope of failover.
- Compliance risk depends on configuration -
data_collection: 'deny'andzdr: trueprovide basic compliance control, but SOC2 and other certifications are not disclosed and need to be confirmed in the business section.
Follow-up observation direction:
- Whether OpenRouter will launch its own model or exclusive hosting model, changing its "pure gateway" positioning
- Continuous optimization of routing algorithms (such as whether multi-objective optimization based on cost + quality will replace pure price routing)
- Compliance certification progress for Enterprise solutions (SOC2, HIPAA, etc.)
- China market/Asia-Pacific node expansion situation
- Depth of third-party integration of MCP Server ecosystem
Related tools: 
GitHub Copilot, 
Cursor
How to use OpenRouter
Get started quickly: API calls
OpenRouter is fully compatible with the OpenAI Chat Completions format, just point baseURL to https://openrouter.ai/api/v1. Here is the minimalist example:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-d '{
"model": "~openai/gpt-latest",
"messages": [
{"role": "user", "content": "Hello, how are you,"}
]
}'Python SDK Example
from openrouter import OpenRouter
import os
client = OpenRouter(api_key=os.getenv("OPENROUTER_API_KEY"))
response = client.chat.send(
model="~openai/gpt-latest",
messages=[{"role": "user", "content": "What is the meaning of life,"}]
)
print(response.choices[0].message.content)TypeScript SDK Example
import { OpenRouter } from '@openrouter/sdk';
const client = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY,
});
const completion = await client.chat.send({
model: '~openai/gpt-latest',
messages: [{ role: 'user', content: 'Hello' }],
});
console.log(completion.choices[0].message.content);MCP Server Configuration
Just add a line of configuration to the AI Coding Assistant to use:
https://mcp.openrouter.ai/mcpAfter logging in via OAuth, the AI assistant can query OpenRouter's model list, price, balance, usage statistics, and documentation in real time.
Comparison table of each entrance
| Entrance | Typical actions | Adaptation team |
|---|---|---|
| API Endpoints | Get API Key → Send Chat Completions Request | Developers, Engineering Team |
| Web console (Settings) | Manage keys, view activity, set budget/policy | Team administrator |
| Chat interface | Different models for online conversation experience | Product managers, reviewers |
| Rankings | Browse model rankings and usage trends | Technology selection decision makers |
| MCP Server | AI coding assistant real-time query model/price/balance | AI application developer |
| Agent SDK | Build AI agents with tool calls | AI agent engineers |
Implementation path: First use the free tier to do API debugging and model comparison for 1-2 scenarios; after confirming the model and provider, switch to the PAYG plan and set provider.order to lock the preferred provider; enable max_price in the production context to prevent accidental overspending, and set data_collection: 'deny' to ensure data compliance.
Version Info
- OpenRouter Online ：The official focus is on continuous iteration of online services, and there is currently no unified public version number and official precise date. Continuously add new models and provider support, and iterate intelligent routing and security capabilities.
- OpenRouter v2 (Growth Stage) ：Introducing Auto Router, Prompt Caching, BYOK, MCP Server and Agent SDK, entering the enterprise-level capability expansion stage.
- OpenRouter Early Online ：There is no official precise date yet, and the milestones visible on the public page are recorded as the minimum version context. In the early days, it focused on aggregating mainstream model providers and establishing basic routing capabilities.
User Reviews