Hall 01 · Free tool · Claude plans

Pick your Claude plan.

Four questions, one answer. Prices straight from Anthropic, the math in plain sight, nothing to sign up for.

Where do you use Claude?
Which plan are you on now?
How often does the 5-hour limit stop you?
And the weekly limit?

Hall 02 · How the picker decides

The math.

Anthropic sells usage per 5-hour session in multiples of Pro. Divide the price by the multiple and the tiers line up. Per week, they line up less neatly.

PlanPer monthUsage per 5-hour sessionPrice per Pro-sized session
Free0€A free cap that varies with demandn/a
Pro18€ (15€ billed yearly)1x, the baseline18€
Max 5x90€5x Pro18€
Max 20x180€20x Pro9€

Max 5x is not a discount. Five times the price buys five times the usage, so a share still costs 18€. You pay for not waiting.

Max 20x is the only tier that gets cheaper, per 5-hour session: 9€ a share instead of 18€. It earns its price once Max 5x stops you most days.

Per week the gap is smaller. Anthropic publishes no weekly numbers for any plan. Classmethod measured three accounts in August 2026 and put the Max 20x weekly limit at about 3.5 times Max 5x, not 4 (range 3.1 to 3.9). If the weekly limit is what stops you, Max 20x saves you about 43% per unit of usage, not half. How much more weekly room Max 5x has over Pro is not published or measured anywhere I could find.

Paid plans carry that weekly limit across all models, and it resets at a fixed time for your account. The picker still treats hitting it every week as a reason to move up one tier, just with smaller expectations.

Claude Code needs a paid plan. On Pro and Max, /usage inside Claude Code shows the same limits as the settings page.

Prices for Germany, excl. VAT, checked 14 October 2026 on claude.com/pricing and in the claude.ai plan settings. Session multiples from Anthropic's Max plan help article. Weekly ratio from Classmethod's measurement (Japanese, 20 August 2026, three accounts, taken during a temporary 50% weekly boost, an estimate, not an official limit). Anthropic publishes no message counts, so this picker uses none.