The Claude API is priced per token, which makes it cheap to start and genuinely surprising at scale. A coding agent that reads a large codebase every turn can burn through tens of millions of input tokens in a day, and at official rates that adds up fast. This guide walks through the actual per-model rates, shows the math for realistic workloads, and gives you a calculator for your own numbers — including what changes when you route Claude through LivRouter's discounted pay-as-you-go plans.
Official Claude API rates
Anthropic prices Claude per million tokens, with separate rates for input, output, and cached input. The catalog snapshot below (September 2026) shows the three tiers most workloads use:
| Model | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|
| claude-sonnet-5 | $3.00 | $15.00 | $0.30 |
| claude-opus-5 | $5.00 | $25.00 | $0.50 |
| claude-haiku-4-5 | $1.00 | $5.00 | $0.10 |
Output tokens cost 5× input across the Claude line, and cached input runs about 10% of the input rate. Prompt caching is the single biggest lever for agent workloads: a system prompt that is re-read on every call gets billed at the cached rate after the first request.
How LivRouter pricing works
LivRouter is a pay-as-you-go gateway: you top up credit ($1 = 10 credits), and every request draws down at a per-model rate derived from the official price. There is no commitment — but there are two optional plans that lower the rates sitewide:
- Free — official model rates, forever, $0/month.
- Pro ($5/mo) — Claude billed at about 30% of official (70% off), plus 60 bonus credits.
- Max ($20/mo) — Claude billed at about 15% of official (85% off), plus 240 bonus credits.
A plan changes the rates and adds bonus credit; the billing model stays pay-as-you-go on every tier. For Claude specifically the per-model effective rates look like this:
| Model | Official (in / out) | Pro (in / out) | Max (in / out) |
|---|---|---|---|
| claude-sonnet-5 | $3.00 / $15.00 | $0.90 / $4.50 | $0.45 / $2.25 |
| claude-opus-5 | $5.00 / $25.00 | $1.50 / $7.50 | $0.75 / $3.75 |
| claude-haiku-4-5 | $1.00 / $5.00 | $0.30 / $1.50 | $0.15 / $0.75 |
Worked example: a support chatbot
Take a customer-facing assistant on claude-sonnet-5 handling a steady month of traffic: 10 million input tokens and 5 million output tokens, no caching. At official rates that is 10 × $3 + 5 × $15 = $105. On the Max plan the same traffic costs 10 × $0.45 + 5 × $2.25 = $15.75 — the $5/month plan fee pays for itself several times over, and the $89.25 difference is budget you keep.
Worked example: a month of Claude Code
Agentic coding is the workload where costs surprise people most, because the agent re-reads project files every turn. A heavy month — 60 million input tokens and 15 million output tokens on claude-sonnet-5 — comes to 60 × $3 + 15 × $15 = $405 at official rates. On the Max plan the identical usage is 60 × $0.45 + 15 × $2.25 = $60.75. With prompt caching enabled on the stable parts of the prompt, the real number drops further, because cached input is billed near the 10% cache rate before the plan discount is even applied.
Interactive cost calculator
- Official cost
- $105.00
- Max ($20/mo) cost
- $15.75
- You keep
- $89.25
Excludes cached input tokens. Rates snapshot 2026-09-07 — live rates in the pricing catalog.
Cached input changes the math
Claude's prompt caching bills repeat prefix tokens at roughly 10% of the input rate. Two things compound in your favor on LivRouter: the plan discount applies to the cached rate too (about 1.5% of official input on Max), and agent frameworks like Claude Code automatically cache their system prompts and file context. If your workload is agent-heavy, model the cache hit ratio before comparing providers — it often matters more than the headline discount.
FAQ
Does the plan discount apply to cached input tokens too?
Yes. Cached input is priced as a fraction of official input rates (about 10% for Claude models), and plan discounts apply on top — so cached input on the Max plan costs roughly 1.5% of the official input rate.
Do I need the Pro or Max plan to get discounted rates?
The Free tier exists forever at official model rates. The optional Pro ($5/mo) and Max ($20/mo) plans lower the rates to about 30% and 15% of official for Claude — all tiers stay pay-as-you-go; a plan changes rates and bonus credit, not the billing model.
Why does output cost 5× input for Claude models?
That ratio comes from the official Anthropic pricing structure that LivRouter's catalog mirrors (a completion ratio of 5). Output tokens cost more everywhere because generation is compute-heavier than reading prompt tokens.
Where are the live rates?
This page uses a snapshot from September 2026. The models & pricing catalog shows live per-model input, output, and cached rates with the official reference price.