What the Claude API really costs

Model prices are quoted per million tokens, separately for input and output. On its own that tells you little: what decides the bill is how you work. Here is how to estimate it and what to watch.

What makes up the bill

Input tokens are everything you sent: the system prompt, the conversation history, file contents, tool results. In agent workloads this is the main expense, because the whole context is resent at every step.

Output tokens are what the model generated. There are usually far fewer of them, but they cost more: Anthropic prices Claude Opus 5.5 output at $20 per million against $4 for input.

A simple consequence follows: chat workloads care about the output price, agent workloads about the input price.

Costing a real session

Take a typical Claude Code run on a mid-sized project. The agent reads files, calls tools, and every step drags the accumulated context along. A dense session runs to roughly 100,000–300,000 input tokens and a few thousand output tokens.

At Anthropic's list rate for Opus 5.5 ($4/$20) that is about $0.40–1.30 per session. A few sessions a day and the month reaches a figure comparable to a subscription, or above it.

Through a gateway such as Zukko↗, where Opus 5.5 costs $0.30 per million on both input and output, the same session costs 3 to 10 cents. One caveat: vendors charge ~10% of the input price for prompt-cache reads, while Zukko bills cached tokens at the regular price, so for agents with a large repeated context the real gap is smaller than a straight comparison suggests.

Subscription or tokens

A subscription suits steady, moderate load: you pay a fixed sum and never count. The downsides are quotas that reset on a schedule, and paying for your peak rather than your actual use.

Tokens suit uneven load and cases where hitting a limit is unacceptable. The downside is that spend needs watching, especially when an agent runs unattended.

Many people keep both: a subscription for daily work, a key for spikes and for the days the quota is gone.

Paying less without losing quality

Match the model to the task. Running Opus on autocomplete is like delivering bread by lorry. Lighter models such as gpt-5.6-luna ($0.10/$0.10) or glm-5.3-flash ($0.04/$0.04) cost less than Opus 5.5 ($0.30/$0.30); whether they handle your routine work — test it on your own tasks.

Watch the context. Every extra thousand tokens in a system prompt is multiplied by the number of agent steps. Trimming context saves more than switching models.

Do not inflate max_tokens without reason. It is a ceiling, but many models tend to fill it.

Look at input price if you work with agents. That is what determines the bill, not the headline output rate.

Current prices

Model prices and limits live in /v1/models/info (or on the Models page in the dashboard); /v1/models returns only the list of ids:

curl https://api.zukko.pro/v1/models/info -H "Authorization: Bearer fetch_your_key"

For comparison, Anthropic's list rates for Opus 5.5 are $4 per million input tokens and $20 per million output. The gap between direct access and a gateway comes from wholesale purchasing terms.

Estimate on your own task rather than someone else's example: consumption depends heavily on codebase size and working style. Top-ups start at $1; there is no starting balance, but subscribing to the @zukkopro↗ channel credits $0.50 once (link Telegram in the dashboard). That is enough to run a couple of real sessions and read your own numbers in the usage history.