This chart covers fourteen priced models from six vendors behind one prepaid credit pool. Because the pricing was never set to be internally coherent, most of the menu is dead weight: nine of the twelve scoreable models are strictly beaten by another model on the same plan that is cheaper on both input and output and at least as intelligent. Not cheaper-but-worse. Just worse. Three models are ever worth picking.
Nothing documents this, but it decodes the entire plan. The Chinese Alibaba site prices these models in
round RMB per million tokens — deepseek-v4-flash is 1 in / 2 out,
deepseek-v4-pro is 12 / 24, glm-5.2 is 8 / 28. A model's credit price is
simply its RMB price times 100. So V4 Flash costs 100 / 200 credits, V4 Pro 1,200 / 2,400.
Measured on the live gateway, not inferred: four models fired with known token counts and their charges read off the bill gave 727.15 / 726.99 / 726.99 / 727.02 credits per dollar — agreeing to 0.02%. That is exactly 100 fen ÷ $0.1375, the fixed rate Alibaba converts RMB at. The dollar prices you see on the international console are the derived view, which is why they are odd numbers (1.65, 3.301, 0.138) while the credit prices are round.
On the buy side, a Max seat turns $200 into 250,000 credits — $344 of usage at list, 1.72x face value. Pro returns 1.38x, Standard just 1.15x. Everything below is priced in credits, because credits are the only thing you actually spend.
Two Qwen models show limited-time discounts on their console pricing pages. The plan honours one and ignores the other. Both were measured directly:
qwen3.7-max, 50% off — honoured. 254,344 input tokens cost 152.55 credits.
Its discounted rate predicts 152.55; its list rate predicts 305.10. You pay 600 credits per 1M, not 1,200.
qwen3.7-plus, 20% off — ignored. 132,732 input tokens cost 26.63 credits.
Its list rate predicts 26.64; its discounted rate predicts 21.30. You pay 200 credits per 1M,
not 160.
Same vendor, same console, same "limited-time" badge, opposite behaviour. There is no rule to infer: whether a discount reaches your credit bill has to be measured per model. Reasoning about credit burn from the console's advertised prices will be wrong for at least one of these two, and you cannot tell which from the page.
Alibaba's terms are explicit. "For interactive use with compatible AI tools only. Not permitted for automated scripts or application backends. Violations may result in subscription suspension or API Key revocation." Keys are per-seat and may not be shared.
So this is a seat subscription for interactive coding tools — Claude Code, OpenClaw, Cursor, Hermes — not cheap inference to build a product on. Every cost figure on this page is bounded by that.
Loaded live from data/models.json, the
machine-readable catalogue this repository maintains against the live gateway. The pricing
analysis further down is a dated snapshot; this table is the current truth about what exists.
| Model | Kind | Tiers | Context | Max out | Released | Notes |
|---|---|---|---|---|---|---|
Loading data/models.json… (needs to be served over HTTP; on a local
file:// open this table stays empty; the JSON is one click away above.)
| ||||||
The three tiers do not buy credits at the same rate. Max is worth 50% more per dollar than Standard.
| Seat | Price / mo | Credits / mo | Credits per $ | API value | Effective | Tokens of v4-flash | Tokens of glm-5.2 |
|---|
Two currencies are in play and it is easy to assume they meet somewhere. They do not. Dollars appear when you buy and on the price display; the charge itself never sees one.
A Max seat. Standard is $30 → 25,000; Pro $100 → 100,000. These are just prices Alibaba set, like tokens at an arcade. Nothing is being converted here.
That last step is not a conversion. A credit is a fen — 0.01 RMB. 1.2 RMB is 120 fen is 120 credits. Same quantity, different name. This is why every model's credit price is a round number: it is just the RMB price with two zeros on the end.
The international console displays the Chinese prices in dollars for your convenience. It is a display layer only. This is why those dollar figures are such odd numbers — $1.65, $3.301, $0.138 — while the credit prices are clean. The odd ones are the derived ones.
You paid $200 for $344 of usage. 1.72x.
This comparison is the only reason a "credits per dollar" number exists at all. It is not a gear in the machine — it is us asking "what would this have cost at list?" so we can judge the seat price. Pro returns 1.38x and Standard 1.15x, so Max is the best-value tier by a wide margin.
Credits per 1M tokens = the model's Chinese RMB price × 100. That is the whole system.
deepseek-v4-flash is 1 RMB in / 2 RMB out, so 100 / 200 credits.
deepseek-v4-pro is 12 / 24 RMB, so 1,200 / 2,400 credits. Everything else on this page is
that fact plus arithmetic.
Everything else is beaten outright at this input size and blend. Move the blend slider: the middle slot swaps between kimi-k2.6 and deepseek-v4-pro as output gets more expensive.
qwen3.7-max holds its place only because of a 50% discount. At its list price of
1,200 / 3,600 credits per 1M it dominates nothing at all, and is itself beaten by
glm-5.2. The
discount is the entire argument for it. qwen3.7-plus is on 20% off. Both are marked
limited-time and neither should be planned around.
Credits per 1M tokens at the selected input size, after any discount. Sorted by intelligence per credit.
| Model | Intel | cr / 1M in | cr / 1M out | cr cached in | Blended | Intel per 1k cr | Context | Max out | RPM |
|---|
deepseek-v4-flash for everything by default. It is 100 / 200 credits per 1M, the
cheapest on the plan by a factor of four, with 1M context, a 384K output ceiling (3x anything else) and
the joint-highest rate limit. It beats six models outright. Its intelligence (40.3) is mid-table, but
nothing cheaper is smarter and nothing near its price comes close.
glm-5.2 when the task is genuinely hard. Highest intelligence on the plan at 51.1,
for 800 / 2,800 credits — roughly 10x V4 Flash. Capped at 500 RPM, so it is not a fan-out model.
qwen3.7-plus when you need what DeepSeek lacks: image and video input, the built-in
Responses-API tools, and 15,000 RPM. It is beaten on pure price-vs-intelligence, but capabilities are
not on this chart and that is a real reason to pick it. Note its tokeniser is hungry (see below), which
erodes its price advantage on the same text.
Ignore deepseek-v4-pro unless your work is output-heavy. It looks like the value
pick on public leaderboards because DeepSeek's own API is dirt cheap. Here it is 12x V4 Flash for four
intelligence points. Its one genuine niche: at 2,400 credits its output is cheaper than
glm-5.2's 2,800, so below about 25% input it re-enters the frontier.
The same 1.1MB of text tokenised to 177,685 tokens on DeepSeek, 201,921 on GLM, and 287,680 on Qwen — 62% more than DeepSeek for identical content. Per-token prices across vendors are therefore not directly comparable, and this chart, like every other, quietly assumes they are. Qwen needs roughly 60% more tokens to say the same thing, which eats most of its headline price advantage.
It has a sharper edge too: that call crossed Qwen's 256k tier boundary and was billed at tier-2 rates on text DeepSeek saw as a 178k prompt. Tier thresholds are counted in the vendor's own tokens, not yours.
500 RPM is the ceiling on every GLM, every Kimi and MiniMax — including glm-5.2, the
smartest model on the plan. The Qwen Plus and Flash models get 15,000, and DeepSeek gets 15,000 on
the mainland sheet. For a wide agent fan-out that gap matters more than price.
Prices: Alibaba Cloud Model Studio console, per-model details pages, Chinese Mainland sheet, captured 16 July 2026, at the rate each model is actually charged (measured, not advertised). Cross-checked against the Chinese site's native RMB sheet: the two are the same prices, converted at 1 RMB = $0.1375, consistent across ten components to 0.36%. Credits at 727 per USD, measured on the live gateway across four models to 0.02%. Alibaba's own docs example implies 724 and is slightly stale; an earlier estimate of ours said 744 and was wrong.
How we know it bills on the mainland sheet. Fit observed credit charges against each sheet and ask whether credits-per-dollar is the same constant for every model. On mainland it is, to within 0.4% across two independent models. On international it varies by 13–27%, so it cannot be the basis. The docs example agrees: against mainland it gives 724, matching our API measurements; against international it gives 400, matching nothing. Four of these fourteen priced models have no international price at all, so it could not bill that way even in principle.
Method. The burn rate was measured, not inferred: ~178k-288k input tokens fired at four models
with thinking disabled and caching defeated, then each call's credit charge read from the console bill.
Implied rates 727.15 / 726.99 / 726.99 / 727.02. Figures here read 100.3 rather than a clean 100 because
the console publishes dollars to only 3 decimals, so converting back adds ~0.3% of noise; the round
number is the real one. Reproduce it with probe-credit-rate.py in this repository.
Seat pricing from the console's Add-seats page, 16 July 2026. "API value" is credits ÷ 727, i.e. what the same usage would cost at list. Token counts assume a 75/25 input:output blend, uncached.
Intelligence: Artificial Analysis Intelligence Index, 19 June 2026. Measured against each
vendor's own API, so it describes the model, not this gateway. qwen3.6-flash and
deepseek-v3.2 have no AA score, so they are listed but not plotted or scored.
Throughput is deliberately absent: AA's tokens/sec is measured elsewhere and would say nothing
true about Model Studio.
Tiers: Qwen models step up above 256k input (3x for qwen3.7-plus, 4x for the 3.6
pair). glm-5 and glm-5.1 step up above just 32k. qwen3.7-max
and all DeepSeek, Kimi and MiniMax models are flat. Models are dropped from a band when their context
cannot reach it.
Caveats. Credits per request also depend on thinking mode and tool calls, neither of which is
modelled here — and reasoning tokens bill as output while ignoring max_tokens, so a
"cheap" call on a reasoning model is not cheap. Built-in tools cost extra on top of tokens
(web_search is $10 per 1,000 calls). Blended cost assumes uncached input.
Full datasheets, the credit-rate experiment, and the metadata errors found along the way:
data/ in this repository.