Claude Code Pricing: What a $200 Max Plan Holds at API Prices

Claude Code comes with every paid Claude plan: Pro at $20 a month, Max 5x at $100 and Max 20x at $200. The alternative is paying per token through the API, where Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Neither price tells you how much work you get, because a plan is two caps, one per five hours and one per week, and Anthropic doesn't say how big either is in tokens or dollars.

I measured them on Max 20x by reading the plan's usage meters through a full window of agent work and pricing that work at API rates. One 5-hour window holds about $300 of Opus 5.5, more than the plan costs for a whole month. A week holds about ten windows, around $3,000. Use all of it every week and that's roughly $13,000 a month of API-priced work for $200.

Claude Code pricing: what each plan includes in October 2026

Every plan runs on two meters. The session meter fills as you work and resets five hours after your first message, and the weekly meter resets at a fixed time each week. Claude Code, the Claude app and the website all draw from both. Anthropic only describes them as multiples of Pro. When you hit one, you wait, or you turn on usage credits and pay API rates for the overflow.

How much work fits in a Max 20x window

Claude Code's API responses carry the plan's live meters in their headers: how full the 5-hour window is and how much of the week is used. I read them through one full window on my Max 20x account while agents worked, then priced the work between readings at API rates from my logs.

On Opus 5.5, 1% of the window was about $3 of API-priced work, so a full window holds roughly $300. The weekly meter moved about a tenth as fast, so a week holds roughly ten windows. That's why I hit the 5-hour wall most weeks and have never hit the weekly one: running out of week would take about one and a half full windows every day.

You can read the same meters on your own account. Any Claude Code request prints them with debug logging on, and a one-line Haiku request costs next to nothing:

ANTHROPIC_LOG=debug claude -p "hi" --model claude-haiku-5-5 --debug 2>&1 \
  | grep -o 'unified-[57][dh]-utilization": "[^"]*"' | sort -u

Both numbers are fractions: 0.43 means 43% used. Read them before and after a stretch of work, price the work's tokens at API rates, and divide by how far the meter moved.

Why one window holds so much: 96% of my tokens were cache reads

An agent doesn't remember anything between steps. Every time Claude Code runs a tool and continues, it sends the whole conversation again: the system prompt, the tool definitions, every file it has read, every command output. Across 305,000 requests in my logs, the average one carried 161,000 tokens of context and produced about 1,100 tokens of output.

Prompt caching is what makes that affordable. The part of the context the model has already processed is billed as a cache read, at a twentieth of the input price on Opus 5.5. Without the cache, the same tokens would have cost about six and a half times as much at API prices.

Cache writes are the part that adds up. Every new file read, tool result or fresh subagent gets written into the cache at 1.25 times the input price, or twice the input price if it's kept for an hour. That's 3% of my tokens and 31% of the cost. Output, the part people usually worry about, was under 1% of the tokens and 23% of the cost. Nearly three quarters of the cost came from subagents, because one session hands out jobs and every subagent builds and pays for its own context. An agent's API bill follows how much context it carries, not how much it writes.

Fable eats the cap twice as fast per dollar

Fable 5.1 costs 2.5 times as much per token as Opus 5.5, so it should drain a plan faster. I wanted to know whether the token price is the whole story. Between two meter readings two minutes apart I ran six Fable 5.1 requests worth $10.94 at API prices. After subtracting the Opus work that ran at the same time, they used about 7.5% of the window.

A dollar of Fable work uses about twice as much of the window as a dollar of Opus work. It's one short test, and with the meter rounding to whole percents the ratio could sit anywhere from 1.8 to 2.4, but older windows agree: three in early September that were mostly Fable ran dry at $147 to $197 of API-priced work. Fable also has its own separate cap on Max. On a plan, that's a strong case for Opus 5.5 as the default and Fable only for the problems that need it. Whether Sonnet 5.5 can be the everyday model instead is a separate question, which I tested in Opus vs Sonnet.

Claude Pro vs Max vs the API: which one I'd pick

The API only makes sense when everything you'd do in a month costs less at API prices than the plan. Agent work gets there fast. A typical session of mine without subagents costs about $1 at API prices, so $20 covers around twenty of them, and the median session overall costs about $6. If your month is a handful of small tasks, pay per token with a spending limit. If you use Claude Code most days, a plan wins by a wide margin: one Max 20x window holds more work than the plan costs for a month. Cursor, which now bills Claude at API list price, lands on the wrong side of that gap; I priced a month of my own work at its rates in Cursor vs Claude Code.

Pro is the way to find out whether Claude Code fits how you work. Its session is a twentieth of Max 20x's by Anthropic's own multiplier, and I've only measured Max 20x, so I won't put a dollar figure on it. Between Max 5x and Max 20x, decide by how often you hit the 5-hour wall. Start on 5x, and move up if the limit stops you more than a couple of times a week.

If you run parallel subagents for hours, take Max 20x and still expect to hit the 5-hour limit. The plan's price stays put while the cap behind it moves whenever Anthropic adjusts its limits, so my $300 is an October number. The meters in the response headers are the only place to see what your own plan holds.