AI token cost calculator for coding agents

How the calculator prices a month

Each kind of token is multiplied by the model's list price per million tokens for that kind, and the four are added up. The team column multiplies by the number of engineers. "Without caching" bills every input-side token (fresh input, cache writes and cache reads) at the fresh input price, as if the provider cached nothing.

The prices are the per-token tables of our Claude Code pricing, Codex pricing and Cursor pricing guides, each with the date it was last checked against the vendor's page. Nothing you type is sent anywhere: the page does the arithmetic in your browser.

Where to find your token counts

• Claude Code: /usage lists input, output, cache read and cache write tokens per model, for the session, the last 24 hours or 7 days. For a month or more, ccusage reads the local transcripts. Our Claude Code token usage guide walks through both.
• Codex: the CLI's /status and /usage show usage by day, by week or in total, and the usage dashboard shows your limits.
• Cursor: grok.com/dashboard and the editor's settings show how much of the weekly pool is used and the prepaid credit left.
• A whole team: each of these reads one person's machine or account. Zest records every engineer's input, output and cache tokens per session, with an estimated cost, next to the pull requests those sessions led to.

Why cache reads are most of a coding agent's bill

A coding agent sends the whole conversation, files and tool results included, on every request. With prompt caching the repeated part is billed as a cache read: on Opus 5.5, $0.20 per million tokens against $4 for fresh input. So a calculator that asks only for input and output tokens prices an agent's month as if nothing were cached.

The month the calculator opens on is one of our engineers': 318 Claude Code sessions (September 4 to October 3, 2026), 67.9B tokens, 98.4% of them cache reads. At Anthropic's list prices on October 4, 2026 it came to $26,072, and the same tokens without caching to $370,526. Where the cost went: cache reads 57%, cache writes 33%, output 10%. That month mixed models and cache durations, and the calculator prices every token on one model, so read the row of the model you run most.

Why a seat price is not the whole cost

Team plans start at a seat price. Claude Team Standard is $25 a seat a month, or $20 billed annually; ChatGPT Business, which includes Codex, $25 per user a month, or $20 billed annually; Cursor Teams Premium $30 per user a month. Each seat includes usage up to limits the vendors don't publish in tokens, and past them usage bills at API rates (Cursor's at API rates plus its token rate). Claude Enterprise is $20 a seat a month, billed annually, with all usage billed at API rates on top. So this calculator gives the price of an API key, of Enterprise usage, or of what a team runs past its seats.

For scale, Anthropic's figure for Claude Code across enterprise deployments is $150–250 per developer a month, about $13 per active day; 90% of users stay under $30 a day. Our measured month is a heavy agent user's, far above that. GitHub Copilot meters in AI credits, not tokens (Copilot Business: $19 a seat, with 1,900 pooled AI credits), so it isn't in the table. The seats side by side are in Claude Code vs Codex vs Cursor.

How the text estimate works

The text box counts characters and words, then estimates tokens with the providers' own rules of thumb for English text. OpenAI puts a token at about 4 characters. Anthropic puts it at about 3.5 characters on earlier models, and says Claude 4.7 and later models, every Claude model priced here, make about 30% more tokens from the same text. Code, other languages and unusual words tokenize differently, so for an exact count use each provider's token counting.