Most developers using Claude Code still hit Anthropic\u2019s API directly. That instinct costs you money and locks you into a single provider\u2019s pricing and availability. The bigger story here isn\u2019t just saving a few cents per million tokens \u2014 it\u2019s about a structural shift in how AI models are accessed.<\/p>\n
The old model was simple: pick one lab, one API key, one bill. The new model is routing. Platforms like OpenRouter aggregate dozens of models from multiple providers under a single endpoint. You get competitive pricing, automatic fallback, and the ability to switch models without rewriting code. For Claude Code specifically, this means you can use Anthropic\u2019s models through a marketplace that often undercuts direct pricing while adding redundancy.<\/p>\n
Here\u2019s how to set it up, step by step.<\/p>\n
Direct Anthropic API pricing for Claude 3.5 Sonnet is $3 per million input tokens and $15 per million output tokens. OpenRouter\u2019s pricing for the same model typically sits around $2.80\u2013$3.00 input and $14.50\u2013$15.00 output \u2014 a small but consistent discount. The real<\/a> savings come from two things: you can cap spending per session (avoid runaway costs), and you can route to cheaper models (like Haiku or Llama) when your task doesn\u2019t need Sonnet\u2019s power.<\/p>\n Why this matters beyond the pennies:<\/strong> OpenRouter lets you treat models as interchangeable commodities. If Anthropic raises prices or suffers an outage, you can shift traffic to another provider serving the same model within minutes. That\u2019s not a feature \u2014 it\u2019s a hedge against vendor lock-in.<\/p>\n Expected outcome:<\/strong> You now have an OpenRouter API key that works with any model they support, including all Claude variants.<\/p>\n Claude Code (the terminal-based agent from Anthropic) supports custom API endpoints via environment variables. OpenRouter exposes a standard OpenAI-compatible endpoint, so the configuration is straightforward.<\/p>\n “`<\/p>\n export ANTHROPIC_BASE_URL=https:\/\/openrouter.ai\/api\/v1<\/p>\n export ANTHROPIC_API_KEY=sk-or-v1-your-key-here<\/p>\n “`<\/p>\n Replace the key with the one you generated.<\/p>\n “`<\/p>\n export CLAUDE_MODEL=anthropic\/claude-3.5-sonnet<\/p>\n “`<\/p>\n OpenRouter uses the format Expected outcome:<\/strong> Claude Code runs identically, but usage is billed through OpenRouter at their rates. You can verify by checking your OpenRouter dashboard for live token counts.<\/p>\n One of OpenRouter\u2019s strongest features is automatic fallback. If the primary model is down or rate-limited, you can define a secondary model to take over.<\/p>\n “`<\/p>\n export CLAUDE_MODEL=anthropic\/claude-3.5-sonnet:google\/gemini-1.5-pro<\/p>\n “`<\/p>\n The colon separates the primary from the fallback.<\/p>\n “`python<\/p>\n import openai<\/p>\n client = openai.OpenAI(<\/p>\n base_url=”https:\/\/openrouter.ai\/api\/v1″,<\/p>\n api_key=”sk-or-v1-your-key-here”<\/p>\n )<\/p>\n response = client.chat.completions.create(<\/p>\n model=”anthropic\/claude-3.5-sonnet”,<\/p>\n fallback_models=[“google\/gemini-1.5-pro”, “meta-llama\/llama-3-70b”],<\/p>\n messages=[{“role”: “user”, “content”: “Write a function to sort a list.”}]<\/p>\n )<\/p>\n “`<\/p>\n Expected outcome:<\/strong> If Claude is unavailable, the request automatically routes to Gemini or Llama without throwing an error. Your code keeps running.<\/p>\n Not all routes are equal. The table below shows typical pricing and latency for the same Claude model through different providers on OpenRouter. Prices are approximate and change \u2014 always check the dashboard.<\/p>\n Tradeoff to watch:<\/strong> Cheaper providers often have higher latency or different output patterns. Test your specific use case before routing everything to the cheapest option.<\/p>\n OpenRouter lets you set per-key spending limits and per-session caps. Do this before running any long automation.<\/p>\n Expected outcome:<\/strong> You stay within budget even if a script loops or a model hallucinates a massive response.<\/p>\n Run a small Claude Code job you already know the answer to \u2014 for example, refactor a Python function. Compare the output from direct API vs. OpenRouter. They should be identical in quality. Check the OpenRouter dashboard to confirm the token count matches.<\/p>\n Why this step is critical:<\/strong> The first run only proves it works once. The second run proves consistency. Run the same request twice and verify the output format matches. If OpenRouter adds any metadata or changes the response shape, adjust your parsing code.<\/p>\n OpenRouter\u2019s real value isn\u2019t the 3% price cut. It\u2019s the ability to treat AI models as a pool rather than a single pipeline. When Anthropic Claude Code\u2019s model is overloaded, you can dip into Gemini or Llama without changing a line of code. When a new model launches, you can test it by swapping one string in your environment variables.<\/p>\n The catch: OpenRouter adds a thin layer of abstraction. If Anthropic changes their API format or deprecates a version, OpenRouter has to update their routing. In practice, they keep pace, but it\u2019s one more thing that can break. For production workloads, keep a direct key as a backup \u2014 not as your daily driver.<\/p>\n","protected":false},"excerpt":{"rendered":" Most developers using Claude Code still hit Anthropic\u2019s API directly. That instinct costs you money and locks you into a single provider\u2019s pricing and availability. The bigger story here isn\u2019t just saving a few cents per million tokens \u2014 it\u2019s about a structural shift in how AI models are accessed. The old model was simple: […]<\/p>\n","protected":false},"author":7,"featured_media":100889,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/96682.png","fifu_image_alt":"Why Your Claude Code API Costs Are Too High (And How OpenRouter","footnotes":""},"categories":[31],"tags":[],"class_list":["post-96682","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology"],"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/96682.png","fifu_image_alt":"Why Your Claude Code API Costs Are Too High (And How OpenRouter","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/96682","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=96682"}],"version-history":[{"count":1,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/96682\/revisions"}],"predecessor-version":[{"id":100890,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/96682\/revisions\/100890"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/100889"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=96682"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=96682"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=96682"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Step 2: Create an OpenRouter Account<\/h2>\n
\n
Step 3: Configure Claude Code to Use OpenRouter<\/h2>\n
\n
\n
provider\/model-name<\/code>codecode for model identifiers.<\/p>\n\n
Step 4: Add Fallback Models for Reliability<\/h2>\n
\n
\n
Step 5: Compare Providers Side by Side<\/h2>\n
\n\n
\n \nProvider<\/th>\n Input Cost (per MTok)<\/th>\n Output Cost (per MTok)<\/th>\n Avg. Latency<\/th>\n Notes<\/th>\n<\/tr>\n<\/thead>\n \n Anthropic (direct)<\/td>\n $3.00<\/td>\n $15.00<\/td>\n ~1.5s<\/td>\n No markup, but no fallback<\/td>\n<\/tr>\n \n OpenRouter (via Anthropic)<\/td>\n $2.90<\/td>\n $14.80<\/td>\n ~1.8s<\/td>\n Slight discount, fallback available<\/td>\n<\/tr>\n \n OpenRouter (via DeepSeek)<\/td>\n $0.14<\/td>\n $0.28<\/td>\n ~2.5s<\/td>\n 20x cheaper, but different model behavior<\/td>\n<\/tr>\n \n OpenRouter (via Google)<\/td>\n $1.25<\/td>\n $5.00<\/td>\n ~1.2s<\/td>\n Gemini 1.5 Pro, good for structured tasks<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Step 6: Automate Spending Controls<\/h2>\n
\n
max_tokens<\/code>codecode parameter to your requests. A typical code generation task might use 4,000 output tokens. Set it explicitly so a runaway agent doesn\u2019t burn through your budget.<\/li>\n<\/ol>\nStep 7: Test with a Real Task<\/h2>\n
The One Thing Most People Miss<\/h2>\n