The top-tier paradox: $200 plans are the cheap ones

Conventional wisdom says start on the $20 tier and upgrade when you hit limits. On Anthropic, Z.ai, and Cursor, the top tier delivers 2–2.25× the tokens per dollar of the entry tier — the small plan is a per-token surcharge you pay for optionality.

At the index’s default assumptions — 25,000 tokens per interaction, 40 active hours a week — Claude Max 20× buys tokens at $0.26 per million effective, exactly half of Pro’s $0.51. Z.ai’s Max tier is 2.25× more token-efficient than its Lite tier. Cursor Ultra is 2.00× more efficient than Cursor Pro. The entry tier is not the cheap seat. On three of the four providers we computed, it is the most expensive way to buy tokens the vendor sells.

This is the opposite of how most developers reason about tiers. The standing advice — start at $20, upgrade when you hit the wall — treats the upgrade as the extravagance. The arithmetic says the surcharge runs the other way: staying small costs you per token, every month, and the discount only unlocks at the top.

Why the paradox exists: quota scales faster than price

When a vendor sells X× the quota for Y× the price, your tokens per dollar move by exactly X ÷ Y — that ratio is the whole story. Anthropic’s pricing page offers “5x or 20x more usage than Pro” at $100 and $200 against Pro’s $20. The 20× tier is twenty times the quota for ten times the price: 20 ÷ 10 = 2.00× the tokens per dollar. No fine print changes that ratio; it falls straight out of the two numbers on the pricing page.

Anthropic: the discount lives only at the very top

Max 20× is twice as token-efficient as Pro, but Max 5× is not more efficient at all — it is a dead heat. Computed from the index’s live dataset: Pro at $20/mo yields 39.0M tokens/mo, 1,950,000 tokens per dollar. Max 5× at $100/mo yields 195.0M tokens/mo — five times the tokens, and an identical 1,950,000 tokens per dollar. The 5× tier is pure headroom at a linear price, per Anthropic’s plan guide (“5x Pro capacity per session” at 5× the cost).

Max 20× at $200/mo breaks the line: 780.0M tokens/mo, 3,900,000 tokens per dollar — exactly 2.00× Pro. The same quota buckets meter Claude Code usage, so the ratio applies to agentic coding work, not just chat. If you buy Anthropic for efficiency rather than headroom, there is exactly one tier that delivers it.

Z.ai: every step up gets cheaper per token

Z.ai is the cleanest version of the pattern — efficiency improves at each tier, not just the last one. The devpack FAQ publishes the quotas: up to ~80 prompts per 5 hours on Lite, ~400 on Pro, ~1,600 on Max. Run through the index at default assumptions (all three tiers bind on their published weekly caps at 40 active hours), Lite at $18/mo computes to 2,407,407 tokens per dollar, Pro at $72/mo to 3,009,259, and Max at $160/mo to 5,416,667 — 2.25× Lite. Top to bottom, that is 20× the weekly quota for 8.9× the price.

Cursor: Ultra’s pool is 20× Pro’s for 10× the price

Cursor states the ratio in dollars, which makes it the easiest to check. Per Cursor’s usage-limits page, Pro at $20/mo includes $20 of API agent usage while Ultra at $200/mo includes $400 — a 20× pool for 10× the price. The index converts those pools at its reference model blend and gets 166,044 tokens per dollar on Pro versus 332,437 on Ultra — 2.00×. The middle tier barely moves the needle — Pro+ at $60/mo includes $70 of usage, only 1.17× Pro’s efficiency. As at Anthropic, the discount concentrates at the top.

Plan	Price/mo	Tokens/mo	Tokens per $	vs entry tier
Claude Pro	$20	39.0M	1,950,000	1.00×
Claude Max 5×	$100	195.0M	1,950,000	1.00×
Claude Max 20×	$200	780.0M	3,900,000	2.00×
Z.ai Lite	$18	43.3M	2,407,407	1.00×
Z.ai Pro	$72	216.7M	3,009,259	1.25×
Z.ai Max	$160	866.7M	5,416,667	2.25×
Cursor Pro	$20	3.3M	166,044	1.00×
Cursor Pro+	$60	11.6M	193,967	1.17×
Cursor Ultra	$200	66.5M	332,437	2.00×

All figures computed by the index at default assumptions (25,000 tokens per interaction, 40 active hours/week); token-native pools converted at the index’s reference blend. Move the sliders and the ratios within a provider barely move — they are set by the vendor’s own quota-to-price arithmetic.

The exception we won’t count: Kimi

Kimi’s marketing supports this thesis harder than anyone’s — and we are not counting it as evidence. The membership pricing page advertises Moderato at $19/mo for 1× Kimi Code credits and Vivace at $199/mo for 30×: a claimed 30× quota for 10.5× the price, which would make Vivace nearly 3× as efficient. But Kimi does not publish exact request counts for the upper tiers, so the index refuses to encode the marketing multiplier and uses conservative estimates instead. Under those estimates the computed pattern inverts: Moderato works out to 2,052,632 tokens per dollar and Vivace to 783,920 — the entry tier looks 2.6× *more* efficient. One of those two pictures is wrong, and until Kimi publishes real quota numbers, we print the conservative one. A pricing page multiplier is a claim; a computed, cap-clipped token count is a measurement.

What the $20 tier actually sells you

The entry tier sells optionality, and optionality has real value — the surcharge is only waste if you could have used the headroom. Effective $/M token counts nothing you don’t consume: Max 20×’s 780M tokens at $0.26/M is a worse deal than Pro’s 39M at $0.51/M for a developer who uses 30M tokens a month, because 750M of that discount evaporates unused. The break-even is consumption. If your usage would saturate the entry tier several times over, the top tier is the cheap plan and “start small” is costing you roughly 2× per token. If it wouldn’t, stay small — you are paying a fair price for the right to walk away.

The standing advice survives in one form: upgrade when you hit limits. Just read the upgrade correctly. Hitting the wall on the $20 tier is not the moment the tool gets expensive — on three of these four providers, it is the moment the discount becomes available to you.

Check the ratios at your own usage shape →