The uncertainty column: which vendors tell you what you’re buying

Alibaba publishes “6,000 requests per 5 hours” — exact. OpenAI publishes a range that spans 5.3×. Anthropic publishes a multiplier of a number it never states. We ranked every vendor’s quota disclosure by how precisely it tells you what you bought.

Coding-plan vendors disclose usage quotas in exactly one of three ways: an exact number (Qwen, MiniMax, Z.ai), a numeric range (OpenAI, Devin), or a bare multiplier of a baseline they never state (Anthropic, Google, Kimi, Factory). The fault line is not price, and it is not geography. The single widest uncertainty band in the dataset belongs to a US agent plan — Devin’s $200 Max tier, at 47–631 messages per day, a 13.4× spread — and the multiplier trick is used identically by a $200 US plan and a $19 Chinese one.

The numbers below come from the sourced low/high bounds the value index maintains for every plan’s native quota. The ranking math uses the point estimate; the bounds themselves have never been surfaced in the app. This is their first publication. The metric is simple: high ÷ low = how much your purchase can vary before the vendor has technically told you anything false. 1.0× means you know what you bought. 13.4× means you don’t.

The three tiers of disclosure

Every plan we track that meters by messages, prompts, or requests falls into one of three disclosure patterns, and the pattern — not the price — predicts how precisely you know your quota.

Exact count. Alibaba’s Qwen coding plan publishes “6,000 requests / 5h · 45,000 / week · 90,000 / month” for its Pro tier. MiniMax and Z.ai commit to single figures too (Z.ai hedges with “approx.” — “Up to approx. 80 prompts” per 5 hours on Lite). Ratio: 1.0×–1.5×.
Numeric range. OpenAI’s Codex pricing page publishes “15–80” GPT-5.5 messages per 5 hours for Plus — a real range, honestly wide. Cognition publishes “8–101 messages/day” for Devin Pro on its Windsurf pricing post. Ratio: 5.3×–13.4×.
Multiplier of an undisclosed base. Anthropic’s plan guide describes Max as “5x Pro capacity per session” — and never states Pro’s capacity. Google and Kimi do the same. Ratio: unknowable from vendor sources alone.

The uncertainty column, ranked

Ranked from least precise to most. Where a vendor publishes a range, the ratio is the vendor’s own. Where a vendor publishes only a multiplier or “dynamic” language, the band is the index’s sourced estimate — marked below — because no vendor number exists to compute one from.

Plan	What the vendor publishes	Band	High ÷ low
Devin Max — $200	“47–631 messages/day” (blog; the pricing page has no numbers)	47–631 / day	13.4×
Devin Pro — $20	“8–101 messages/day”	8–101 / day	12.6×
OpenAI Plus / Pro 5× / Pro 20×	“15–80”, “75–400”, “300–1600” msgs / 5h	vendor range	5.3× (all tiers)
Kimi Allegretto–Vivace — $39–199	“5x / 15x / 30x” quota, no absolute number	300–1,200 / 5h — our estimate	4.0×*
Anthropic Pro / Max — $20–200	“dynamic” 5h limits; Max is “5x/20x Pro capacity”	30–70 msgs / 5h (Pro) — our estimate	2.3×*
Google AI Pro / Ultra — $20–200	Ultra is a “5X/20X higher usage limit” than Pro; Pro has no number	30–70 / 5h (Pro) — our estimate	2.3×*
Factory Droid tiers — $20–200	prices, multipliers, limit structure — no message counts	30–70 / 5h (Pro) — our estimate	2.3×*
xAI SuperGrok / Heavy	prices only; no official message windows	100–150 / 4h — our estimate	1.5×*
Z.ai Lite / Pro / Max — $6–60	“Up to approx. 80 / 400 / 1,600 prompts” per 5h	single figure + “approx.”	1.0–1.5×
Qwen Lite / Pro — $8–40	“1,200” / “6,000 requests / 5h” + weekly and monthly caps	exact	1.0×
MiniMax Plus / Max / Ultra	single request-equivalent figures, rolling + weekly usage bar	exact	1.0×

*Starred ratios are the index’s own secondary estimates, not vendor disclosures. That distinction matters and cuts the uncomfortable way: OpenAI’s 5.3× looks worse than Anthropic’s 2.3× in the table, but OpenAI at least published a range you can hold it to. Anthropic, Google, Kimi’s upper tiers, and Factory published nothing — the 2.3× band exists only because we built one.

The widest band belongs to Devin, not OpenAI

If you expected a marquee lab to own the vaguest quota, the data says otherwise: Cognition’s Devin Max spans 47–631 messages per day — a 13.4× spread, per its own Windsurf pricing post, more than double OpenAI’s widest ratio. The main devin.ai pricing page gives no numbers at all — just an assurance that the allowance “will be enough to fully cover all agent usage.” A 13× range on a $200/month plan means the same dollar buys you a light day or a heavy fortnight of work, and the vendor considers both outcomes as-advertised.

Paying OpenAI 10× more buys zero extra precision

Within OpenAI’s own ladder, the uncertainty ratio is invariant: Plus at $20 is “15–80” messages per 5 hours, Pro 5× at $100 is “75–400”, Pro 20× at $200 is “300–1600” — exactly 5.33× on all three tiers. Ten times the money buys ten times the capacity and not one bit more certainty about what that capacity is. The range scales; the fog is constant.

The multiplier trick cuts across geography

The lazy version of this story — opaque US premiums, precise cheap Chinese plans — does not survive contact with Kimi. Moonshot’s own pricing page discloses Kimi Code quota only as multipliers — 1x, 5x, 15x, 30x across four tiers from $19 to $199 — with no absolute count for the three upper tiers. That is structurally identical to Anthropic’s “5x Pro capacity per session” and Google’s “5X higher usage limit… than our Pro plan” — in each case a multiplication against a number the vendor declines to state. Meanwhile Alibaba publishes exact per-window, per-week, and per-month request counts, plus a conversion note that one request covers roughly 5–10 model calls for simple tasks and 10–30+ for complex ones. Precision is a disclosure choice, not a regional trait.

What the exact-count vendors prove

Publishing a real number is clearly possible — Qwen, MiniMax, and Z.ai run the same class of metered inference business and manage it. The usual defense of “dynamic” limits is that load varies and a hard number would be a hostage to peak demand. Z.ai’s answer is instructive: one figure per tier with an “approx.” qualifier. That single word does everything “dynamic limits” claims to do, while still telling the buyer what they bought. Credit-pool tools (Cursor, Copilot, and the app-builders) sit outside this ranking entirely — they publish exact credit counts but hide the credit-to-work conversion instead, which is a different opacity worth its own piece.

How to read the column

A published range beats a published multiplier, and an exact count beats both — because a range at least bounds your downside. If you buy OpenAI Plus, the worst case is written down: 15 messages per 5 hours. If you buy Claude Max 5×, the worst case is 5 times a number that does not exist. The index prices every plan at a sourced point estimate inside these bands, and the assumptions are adjustable precisely so you can stress-test the vague ones. But the column itself is the finding: five of the ten vendor families above sell a quota they decline to quantify.

See how we turn these bands into $/M token →