The frontier tax: what you actually pay for the last 20% of capability

MiniMax delivers a million tokens for a nickel. Cursor’s frontier pool charges $6 for the same million. The capability gap between them is real — but is it 120× real?

Normalized to tokens per dollar, the premium for a frontier-class coding plan runs from 10× (Claude Pro vs MiniMax Plus) to 117× (Cursor Pro’s credit pool vs the same MiniMax tier). The measured capability gap between the models those plans run — Claude Opus 4.8 vs MiniMax-M3 on the Artificial Analysis coding index — is 1.31×. You pay ten to a hundred times more for 31% more measured capability. That mismatch is the frontier tax, and it has a defensible number.

The numbers below come from the live value index at its default assumptions — 25,000 tokens per interaction, 40 active hours per week — which normalizes every plan’s native quota (messages, requests, credits) into estimated monthly tokens and computes an effective $/M token. Move the sliders and the exact figures shift; the shape of the gap does not.

A nickel versus six dollars

MiniMax Plus costs $20/month and computes to $0.0513 per million tokens; Cursor Pro costs the same $20 and computes to $6.02 per million — a 117.4× spread between two plans with identical sticker prices. MiniMax’s Plus tier grants 4,500 request-equivalents per 5-hour window, which at the default assumptions works out to roughly 390M tokens a month. Cursor Pro grants a monthly API-usage pool worth about $20 of frontier-model calls, which the index converts to ~3.3M tokens.

The dek’s “$6 versus a nickel” rounds to an even 120×; the precise figure at default assumptions is 117.4×. Both are honest, but 117× is the computed one. One methodological note: MiniMax’s number traces to a native per-plan quota, while Cursor sells a dollar-denominated pool with no published per-token rate — the index converts it using a documented frontier-blend reference rate ($3 in / $15 out, per Cursor’s own model pricing). The $6.02 is the index’s estimate of Cursor’s pool, not a rate Cursor states.

The capability gap is 31%, not 100×

On the Artificial Analysis coding index, Claude Opus 4.8 scores 56.7 and MiniMax-M3 scores 43.4 — a 1.31× gap. The intelligence index tells the same story: 56 versus 44, or 1.27×. Against that, Claude Pro at $20/month with ~45 messages per 5-hour window computes to $0.5128/Mtok — exactly 10.00× MiniMax Plus. Cursor Pro and GitHub Copilot Business, both credit pools priced off frontier API rates, land at 117×.

Plan	Price	Effective $/Mtok	Model (AA coding index)	Premium vs MiniMax Plus
MiniMax Max	$50/mo	$0.0385	MiniMax-M3 (43.4)	0.75×
MiniMax Plus	$20/mo	$0.0513	MiniMax-M3 (43.4)	1× (baseline)
Claude Pro	$20/mo	$0.5128	Claude Opus 4.8 (56.7)	10.0×
ChatGPT Plus	$20/mo	$0.5128	GPT-5.5 (74.9)	10.0×
GitHub Copilot Pro	$10/mo	$4.01	frontier credit pool*	78×
GitHub Copilot Business	$19/mo	$6.01	frontier credit pool*	117×
Cursor Pro	$20/mo	$6.02	frontier credit pool*	117×

*Credit-pool rows are the index’s reference-blend estimates of dollar-denominated pools, not vendor-published per-token rates. And a necessary caveat on the other axis: the AA coding index is one composite benchmark, not a complete stand-in for how a model performs on your stack. The argument here is a value argument on that specific metric — a widely used, independently maintained one, but a metric all the same.

Not every frontier premium is equally steep

The same 10× premium buys different amounts of capability depending on whose frontier you buy. GPT-5.5 posts 74.9 on the coding index — the highest score in the dataset, 1.73× MiniMax-M3 — and ChatGPT Plus computes to the same $0.5128/Mtok as Claude Pro. So OpenAI’s 10× buys 73% more measured capability where Claude Pro’s 10× buys 31%. The frontier tax is least defensible not at the frontier labs’ own $20 tiers, but at the IDE credit pools charging 117× for access to the same models the labs sell at 10×.

Weighted for capability, the spread is still 120×

Even after crediting frontier models for their higher index, the gap barely narrows. The index’s coding-value metric — tokens per dollar weighted by the model’s AA coding score — puts MiniMax Max at the top of all 64 plans (11.28M) and Cursor Pro 47th (94,147): a 119.86× spread with capability already priced in. GLM-5.2’s $160 Max plan ranks fourth at 3.73M, behind three MiniMax tiers; the seventeen plans that score below Cursor Pro are almost all other credit pools, agent tools, and app-builder tiers. If the capability weighting were doing the work the premium implies, the spread would collapse toward 1×. It stays at 120×.

This is not one dataset’s quirk. ThinkAI’s analysis of the collapsing frontier moat finds frontier proprietary models running 3–5% ahead of open alternatives on standard benchmarks while costing 5–10× more at volume. Their multiples are smaller than ours — they measure raw API rates, not subscription quotas — but the direction is identical: price spreads dwarfing capability spreads.

When the tax is rational

Pay the frontier tax when the task is capability-bound; skip it when the task is volume-bound. A bug the cheaper model cannot fix costs you infinite tokens at any price, and for that work the 10× — even the 117× — is trivially worth it. The economics only break when you pay the frontier rate for work that isn’t capability-bound: boilerplate, test scaffolding, mechanical refactors, glue code. That is most daily coding, and at 117× the break is expensive.

If your work is mostly high-volume and routine, a $0.05/Mtok-class plan covers it at 1/10th to 1/117th the effective rate.
If you need frontier capability daily, buy it from a lab’s native plan (10×) rather than an IDE credit pool (78–117×) — the model is the same Opus 4.8; the markup is not.
If you need it occasionally, pay the tax per task, not per month: default cheap, escalate hard problems.

These are estimates from adjustable assumptions, not vendor guarantees — the index exists so you can compute the tax at your own usage shape rather than ours.

Compute the frontier tax at your usage shape →