Six plans cost $20. Their token allowances differ by 120×.

Claude Pro, ChatGPT Plus, Cursor Pro, Copilot Business, Google AI Pro, and MiniMax Plus all sell within a dollar of $20 a month. Normalized to tokens, the effective price of a million tokens runs from $0.05 to $6.02 — the same sticker buys wildly different products.

Six coding plans bill within a dollar of $20 a month: Claude Pro at $20, ChatGPT Plus at $20, Google AI Pro at $19.99, MiniMax Plus at $20, Cursor Pro at $20, and Copilot Business at $19 per seat. Run all six through the index’s normalization at its default assumptions and the effective price of a million tokens spans $0.05 (MiniMax Plus) to $6.02 (Cursor Pro) — a 117× spread. The sticker price, the one number every pricing page leads with, is the least informative number on any of them.

None of these vendors publishes its allowance in comparable units — one counts messages per five-hour window, one sells requests, two sell dollar-denominated credit pools. The methodology converts each plan’s native unit to estimated tokens per month, then divides price by tokens. What follows is that computation for the ~$20 tier, side by side.

What $20 actually buys

At the index’s defaults — 25,000 tokens per interaction, 40 active hours a week — the same ~$20 buys anywhere from ~3.2 million to ~390 million tokens a month. Every figure below re-computes live as you change those assumptions on the compare view.

Plan	Sticker	Est. tokens/month	Effective $/M	API-equivalent multiple
MiniMax Plus	$20	~390M	$0.05	6.9×
Google AI Pro	$19.99	~42M	$0.48	5.5×
Claude Pro	$20	~39M	$0.51	12.2×
ChatGPT Plus	$20	~39M	$0.51	12.8×
Copilot Business	$19/seat	~3.2M	$6.01	0.7×
Cursor Pro	$20	~3.3M	$6.02	0.6×

Two ratios sit behind the “120×” in the title, and both round to it: token allowances span 122× (3.2M to 390M a month) and effective $/M spans 117.4×. They differ because the plans at the extremes differ — MiniMax against Cursor on price-per-token, MiniMax against Copilot on raw tokens. Either way the spread moves as your usage shape moves, which is why the index makes it adjustable rather than printing one “best value” verdict.

Why MiniMax lands at $0.05

MiniMax Plus is roughly ten times cheaper per token than Claude Pro or ChatGPT Plus because its quota is shaped differently: the index’s comparator estimate is ~4,500 requests per five-hour window, against ~45 messages for the incumbents. MiniMax’s own plan page describes the quota as rolling five-hour and weekly windows rather than a fixed count, so the index scales each “request” to a tenth of a full interaction — and the plan still computes to ~390M tokens a month, an order of magnitude above the ~39M that Claude Pro and ChatGPT Plus yield at the same sticker.

The cheapness is real, but it buys MiniMax-M3, not a frontier Western model. Whether that trade is worth 10× the headroom is a capability question the cost lens alone can’t answer — it is the reason the index ranks by capability-adjusted value as well as raw $/M.

The expensive end is a mechanism, not a vendor

Cursor Pro ($6.02/M) and Copilot Business ($6.01/M) are effectively tied at the top of the price range, and for the same structural reason: neither is a quota plan. Cursor Pro includes $20 of API usage credit for its Auto and Composer pool, metered at per-model token rates; Copilot meters AI Credits against a per-model rate card. Both are “buy $20 of API credit, redeem at market model rates.” Convert a $20 pool at the index’s ~$6-per-million-token reference blend and you get ~3.3M tokens — the plans land at $6/M nearly by construction.

That reframes the usual complaint. Cursor is not uniquely expensive and GitHub is not uniquely stingy; a dollar-denominated pool simply passes API list pricing through to you, while the quota plans above them in the table are selling tokens below what the same usage would cost at the API counter.

The Copilot asterisk

GitHub does not sell an individual $20 plan. Its retail tiers are Pro at $10, Pro+ at $39, and Max at $100; the only ~$20 Copilot price point is Business at $19 per granted seat, an org-pooled, per-seat SKU you buy for a team, not for yourself. It is in this face-off because it is the closest-priced Copilot plan, but it is a different purchase shape than the five personal subscriptions beside it — a seam worth naming rather than papering over.

What the API-equivalent multiple separates

The same six plans split cleanly into subsidized and unsubsidized. At default assumptions, four plans deliver several times their price in API-equivalent usage: ChatGPT Plus at 12.8×, Claude Pro at 12.2×, MiniMax Plus at 6.9× (against MiniMax’s pay-as-you-go rates), and Google AI Pro at 5.5×. The two credit pools sit below parity — Cursor Pro at 0.62×, Copilot Business at 0.70× — meaning the included usage costs about what the raw API would, minus the pool’s overhead.

Read as a buying rule: if your workload is token-hungry and model-flexible, the quota subscriptions are underpriced relative to the API and the pools are not. The pools buy something else — routing, IDE integration, org billing — and should be judged on that, not on tokens per dollar.

Where these numbers bend

Several of the native limits feeding this comparison are the index’s comparator estimates, not official fixed counts. Anthropic’s ~45 messages per five-hour window, Google’s ~48 credits, and MiniMax’s ~4,500 requests are point estimates inside published ranges; the vendors themselves publish dynamic usage language — Anthropic describes Pro limits qualitatively, and OpenAI publishes a 15–80 message range for the Plus five-hour window. The index carries low/high bands and a per-plan confidence field for exactly this reason, and the dataset is a dated snapshot (2026-06-19), not a live vendor feed.

The spread itself is robust to all of that. Move every estimate to the generous end of its band and the ~$20 tier still spans two orders of magnitude in effective $/M — which is the finding: the tier you compare on is the one number that tells you nothing.

Re-rank the ~$20 tier on your own usage →