A “request” is not a prompt: how quota units inflate by 5–30×

Alibaba sells its Qwen coding Pro plan on a big number: 6,000 requests per five hours, per its own pricing page. The same page then defines the unit away: “Each query consumes quota based on the number of model calls. Simple tasks typically use 5–10 calls, while complex tasks may use 10–30 or more.” A “request” is one backend API call. The thing you experience — typing a prompt and waiting for the agent to finish — burns five to thirty of them. So 6,000 requests is somewhere between 200 and 1,200 actual interactions, and the headline overstates usable work by 5–30×, on the vendor’s own numbers.

This matters because every tokens-per-dollar comparison starts from the quota. Take the 6,000 at face value and Qwen Pro computes to about a penny per million tokens — an order of magnitude cheaper than it plausibly is. Any ranking that skips the deflation step is comparing fiction. The value index applies a per-plan correction before any dollar math runs; this piece shows the mechanism and the before/after for each request-metered plan.

The vendor said it out loud

Alibaba is the only request-quota vendor that publishes its own inflation ratio. The coding plan page states it verbatim: “Simple tasks typically use 5–10 calls, while complex tasks may use 10–30 or more.” That sentence sits directly under quota tables listing 6,000 requests per 5 hours, 45,000 per week, and 90,000 per month for Pro — exact figures, honestly disclosed, in a unit no buyer would guess means “backend call.” The coding plan FAQ adds the enforcement: separate quota-exceeded errors for the 5-hour window, the week (reset Monday 00:00 UTC+8), and the month. Three walls, all denominated in the inflated unit.

The gap between the unit and the experience is not hypothetical. In qwen-code issue #3267, a free-tier developer reports exhausting a 1,000-requests-per-day allowance “without completing even one task today,” and asks for any visibility into what “managed to consume 1,000 requests in an hour.” One thousand of something per day sounds generous. Divided by 5–30, it can be a single unfinished afternoon.

The correction: deflate before you divide

The index never takes a request count at face value. Each request-metered plan carries a tokensPerUnitScale — 0.1 for Qwen and MiniMax, 0.15 for Kimi — that maps backend requests down to prompt-interaction equivalents before any tokens or dollars are computed. At the default assumption of 25,000 tokens per interaction, a Qwen “request” is therefore worth 2,500 tokens, not 25,000. The 0.1 is not a round number pulled from the air: a 10× deflation sits inside Alibaba’s own published 5–30× range, toward the conservative end.

Worked through for Qwen Pro at the index’s defaults (25,000 tokens per interaction, 40 active hours per week): 6,000 requests × 2,500 tokens is 15M tokens per 5-hour window. Eight windows a week would be 48,000 requests, which the published 45,000 weekly cap clips first — so the plan yields about 112.5M tokens per week, 487.5M per month, and $50 ÷ 487.5M ≈ $0.10 per million tokens. Skip the deflation and the same arithmetic says $0.0103 — a plan that looks exactly 10× cheaper than the corrected figure.

All rows computed at the index defaults above. The Qwen scale is anchored to Alibaba’s published ratio; the Kimi and MiniMax scales are the index’s comparator estimates, flagged as such in the dataset — and MiniMax’s per-window request count itself is an index estimate, because its official pricing docs publish no fixed per-5-hour figure at all.

Kimi and MiniMax: the same myth, without the confession

Neither Kimi nor MiniMax publishes a calls-per-query ratio — what they publish instead is a unit nobody can price. Moonshot’s membership pricing page prices four tiers from $19 to $199 but expresses Kimi Code quota only as relative multipliers — 1×, 5×, 15×, 30× — with no absolute request count for any tier above the base. MiniMax’s Token Plan FAQ describes “usage-based deduction” in three complexity bands (chat is cheap, code generation costs more, long-context agent work costs most) and never states a numeric ratio. That is a different mechanic from Qwen’s — opacity about what a unit is worth, rather than a stated inflation — but it breaks naive comparison the same way: the headline count is not denominated in anything you can experience.

Kimi has since taken the logical next step. Its membership credits explainer says the whole membership line now runs on a unified token-based credit pool — “Credit consumption is based on token usage,” with writing a piece of code pegged at roughly 0.5–2% of credits. That is a recent shift, and it postdates the dataset’s June verification of the coding tiers. But read it as the vendor conceding the premise: fixed-count units never mapped cleanly to work, so Kimi is abandoning them for token-weighted ones.

How to deflate any quota unit

Before comparing any two request-metered plans, normalize the unit — the ranking is meaningless otherwise. The procedure is short:

• Find the unit’s definition, not its count. It is usually buried under “quota consumption” or “deduction rules,” below the pricing table. If the word is “request” or “call,” assume it is smaller than a prompt until proven otherwise.
• Apply the vendor’s ratio if one exists. Alibaba’s 5–30 calls per query means dividing the headline by 10 is generous to the vendor.
• Treat unratioed units as an upper bound. Kimi’s multipliers and MiniMax’s complexity bands give you no divisor — so any point estimate (including the index’s 0.15 and 0.1) is a labeled assumption, not a fact. Adjust it and watch whether your ranking survives.

The liftable version: at identical face-value arithmetic, Qwen Pro reads as $0.0103 per million tokens; corrected by Alibaba’s own 5–30 calls-per-query disclosure, it is about $0.10 — still cheap, but ten times less miraculous. The index bakes this correction into every request-metered plan and exposes the assumption as a slider, because a number you can stress-test beats a number you have to believe.

See the full normalization pipeline →

Sources

• Alibaba Model Studio — coding plan (quota tables + 5–30 calls-per-query note)
• Alibaba Model Studio — coding plan FAQ (5-hour / weekly / monthly quota errors)
• GitHub — QwenLM/qwen-code issue #3267 (1,000 requests/day, zero tasks finished)
• Moonshot — Kimi membership pricing (1×/5×/15×/30× quota multipliers)
• Moonshot — Kimi membership credits (token-based credit consumption)
• MiniMax — Token Plan FAQ (usage-based deduction, three complexity bands)
• MiniMax — Token Plan pricing (no fixed per-5h request counts)
• theverdict.ai — methodology