Pick a model, set your agent's usage (including prompt caching), and see the per-token bill — live OpenRouter pricing — against Standard Compute's flat plan.
Monthly per-token cost (input/output + caching), cheapest first. Your selected model is highlighted.
Public OpenRouter per-token pricing (input + output, with your cache-hit rate applied where the model supports prompt caching). Standard Compute is flat with no usage limits within fair use; suggested plan scales with workload.
Point any OpenAI-compatible agent at Standard Compute and the bill stops scaling with usage. One flat price, no usage limits, smart routing to a cost-efficient frontier model per request.
Every major LLM API — OpenAI, Anthropic (Claude), Google (Gemini), xAI (Grok) — bills per token. You pay for input tokens (your prompt, context, files) plus output tokens (what the model generates), each priced per million. Output usually costs 3–5× more than input. Many models also offer prompt caching, where repeated context is billed at a fraction of the input price.
That model punishes AI agents: they resend large context every step and fire background requests, so a single always-on agent can burn millions of tokens a day. This calculator estimates that monthly per-token bill across 30+ frontier models (live OpenRouter pricing, caching included) and compares it to a flat-rate plan with no usage limits where cost stops scaling with usage.
LLM API cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price), multiplied by your number of requests. Input (prompt) tokens and output (completion) tokens are usually priced differently, and output is typically 3–5× more expensive per token. This calculator does the math across 30+ models using live OpenRouter pricing.
Per-token API pricing varies widely by model — from cents per million tokens for small open models to $15+ per million output tokens for frontier models like GPT-5.5 and Claude Opus. Pick any model above to see its exact per-token cost at your usage, or compare several side by side.
Agents are input-token heavy: they resend large context (files, history, tool results) on every step, and background heartbeats, retries, and parallel tool calls multiply the request count. A single always-on agent can push millions of tokens a day, so per-token bills scale fast — which is why flat-rate pricing changes the economics for agent workloads.
Input (prompt) tokens are what you send to the model — your prompt, context, files, and message history. Output (completion) tokens are what the model generates back. Output tokens usually cost several times more than input tokens, so a chatty agent with long responses can be pricier than the request count suggests.
Yes — Standard Compute is an OpenAI-compatible API with no usage limits at a flat monthly price (from $39), instead of per-token billing. It auto-routes each request to a cost-efficient frontier model and, being flat-rate, adapts response speed under heavier load rather than charging more or throwing rate limits. That's what the comparison above shows.
Yes — a lot, for agents. Many models bill repeated input (system prompt, context, files you resend every step) at a discounted cache-read rate, often around 10% of the normal input price. Since agents resend large context constantly, a high cache-hit rate can cut the input portion of the bill dramatically. This calculator lets you set your cache-hit rate and applies the model's cache-read price where the model supports it.
It depends on context size and how often it runs, but agent inputs commonly range from ~10,000 tokens up to 1,000,000+ per request for large-context or codebase-aware agents, with outputs from a few hundred to a few thousand tokens. Set your real numbers on the sliders above to estimate the bill.