Three things to remember
- Eight frontier points run from ¥0.5 to ¥18; the other 18 can each be matched by a cheaper model with no lower composite.
- Three cliffs: ¥0.6→2.5 (4.2× for 1.3), ¥2.5→6 (2.4× for 0.9) and ¥12→15 (1.25× for 0.6); the last ¥3 buys 0.2 points.
- Two empty bands: ¥2.5–4 and ¥12–15. Price and score correlate at 0.44, explaining only 19% of variance.
The price table is not a curve
Sort the 26 models with public prices by output price and you get a strange curve: not continuous, but a staircase.
- Cheapest ¥0.5 (MiniCPM-V 4.5), most expensive ¥20 (Pangu 5.5);
- 26 models use only 20 distinct price points; the median is ¥6;
- six models are at ¥0 (see zero-price models) and sit outside this curve.
Put price and composite together and most models are not on the “price-performance frontier”. Only eight of 26 are on it; the other 18 can each be matched or beaten by a cheaper model. That does not make the 18 useless — they may sell modalities, licences or industry solutions — but if you buy purely on “score per yuan”, they are not the answer.
This essay does three things: find the eight frontier points, mark three price cliffs, and identify two price bands with no model. Every number can be recomputed from the tier list and model library.
The eight frontier points
Frontier definition: sort the 26 by price ascending and keep only points whose composite is higher than every cheaper model. Each survivor is “at this price, you cannot buy a higher score”.
| Price / 1M output tokens | Model | Composite | Overall | Versus previous |
|---|---|---|---|---|
| ¥0.5 | MiniCPM-V 4.5 | 79.6 | #26 | — |
| ¥0.6 | DeepSeek-V4.1-Flash | 87.9 | #12 | +8.3 pts / +¥0.1 |
| ¥2.5 | LongCat-Next | 89.2 | #9 | +1.3 pts / +¥1.9 |
| ¥6 | MiMo-V2.5 | 90.1 | #7 | +0.9 pts / +¥3.5 |
| ¥9 | GLM-5.3 | 92.4 | #4 | +2.3 pts / +¥3 |
| ¥12 | Qwen3.5-397B | 94.6 | #3 | +2.2 pts / +¥3 |
| ¥15 | Doubao Seed 2.0 Pro | 95.2 | #2 | +0.6 pts / +¥3 |
| ¥18 | Qwen3.8-Max | 95.4 | #1 | +0.2 pts / +¥3 |
The first row is the most counterintuitive: from ¥0.5 to ¥0.6, ¥0.1 more buys 8.3 points. This is not “price buys performance”; MiniCPM is an 8B edge model and DeepSeek-V4.1-Flash is a different product class. The ¥0.1 crosses two categories.
The last row is equally counterintuitive: from ¥15 to ¥18, ¥3 more buys 0.2 points. And that 0.2 sits inside grade-B error bars (see the opacity index).
The frontier is not a straight line; it is steep then flat. In the cheap tier each yuan buys many points; in the expensive tier each yuan buys few — but the cheap tier has a tiny denominator and the expensive tier has a higher absolute score, so both ends have buyers.
Three cliffs
Put the price increase and score increase of adjacent frontier points side by side:
| Cliff | Price change | Increase | Score change | Points per +¥1 |
|---|---|---|---|---|
| One: ¥0.6 → ¥2.5 | ×4.2 | +317% | +1.3 | 0.68 |
| Two: ¥2.5 → ¥6 | ×2.4 | +140% | +0.9 | 0.26 |
| Three: ¥12 → ¥15 | ×1.25 | +25% | +0.6 | 0.20 |
Cliff one: the end of the cheap tier. DeepSeek-V4.1-Flash at ¥0.6 already scores 87.9; to reach 89.2 you jump to ¥2.5 (LongCat-Next), 4.2× the price. The 1.3 points buy “native multimodality plus generation 92”, not “cheaper documents”.
Cliff two: from specialist to generalist. ¥2.5 to ¥6, 2.4× price for +0.9 points. This segment buys generality, not score: MiMo has video 93 and ecosystem 88; LongCat has ecosystem 72. If your scenario does not need video, you can skip this premium.
Cliff three: flagship premium. Qwen3.5 at ¥12 already scores 94.6; Doubao at ¥15 scores 95.2; Qwen3.8-Max at ¥18 scores 95.4. From ¥12 to ¥18, price +50%, score +0.8. Of those 0.8, 0.2 sits inside grade-B error; what you actually buy is the #1 position, plus Qwen3.8-Max’s three 95+ scores in perception, chart and agent.
Marginal return: the last ¥3 buys 0.2 points
Compute “points per extra ¥1” separately:
| Interval | Points per +¥1 |
|---|---|
| ¥0.5 → ¥0.6 | 83.0 |
| ¥0.6 → ¥2.5 | 0.68 |
| ¥2.5 → ¥6 | 0.26 |
| ¥6 → ¥9 | 0.77 |
| ¥9 → ¥12 | 0.73 |
| ¥12 → ¥15 | 0.20 |
| ¥15 → ¥18 | 0.07 |
Two high-return intervals: ¥6→¥9 (+2.3 points, +50%) and ¥9→¥12 (+2.2 points, +33%). These are the last value zones: GLM-5.3 pushes agents to 95; Qwen3.5 pushes document/chart/ecosystem to 96–98.
Beyond ¥12, marginal return collapses: ¥12→¥15 is 0.20 points per yuan; ¥15→¥18 is 0.07. If your budget is above ¥12, you are no longer buying “score”; you are buying “rank #1, modality, ecosystem or compliance”. Not that flagships are bad — Qwen3.8-Max’s perception 97, chart 96 and agent 95 are real capability — but the reason for the flagship premium must come from outside the score.
Two empty bands
Among the 26 price points, two bands are empty:
Nothing between ¥2.5 and ¥4. Above is Qwen3-Omni-30B at ¥4 (88.9, all-modality); below is LongCat-Next at ¥2.5 (89.2, native multimodal). The empty band means: at a budget of about ¥3, you cannot buy “stronger than LongCat, cheaper than Omni”. Either pay up for all-modality or stay at ¥2.5.
Nothing between ¥12 and ¥15. Above is Doubao at ¥15 (95.2); below is Qwen3.5 at ¥12 (94.6). This gap is more expensive: you cannot spend ¥13 to buy “0.3 points above Qwen3.5, ¥2 below Doubao”. Either stop at Qwen3.5 or jump to Doubao.
The widths differ; the meaning is the same: Chinese multimodal pricing is not continuous competition, but a few vendors anchoring a few price points. A new entrant avoiding a head-on fight has two natural slots at ¥3.5 and ¥13.5 — an observation, not advice; pricing still depends on cost and scenario.
Price and score: r = 0.44, 19% explained
Across the 26 priced models, Pearson r between price and composite is 0.44, R² = 0.19. Price explains only 19% of score variance; 81% comes from elsewhere.
By band:
| Band | Models | Average score | Average rank | Examples |
|---|---|---|---|---|
| ¥0–1.5 | 5 | 84.4 | 17.4 | DeepSeek-V4.1-Flash, Qwen3.8-Flash |
| ¥1.5–4 | 2 | 87.6 | 11.5 | LongCat-Next, Keye-VL-2.0 |
| ¥4–10 | 12 | 84.2 | 17.7 | GLM-5.3, MiMo, Omni, ERNIE |
| ¥10–15 | 3 | 92.2 | 5.7 | Qwen3.5, GLM-5, SenseNova |
| ¥15–21 | 4 | 91.1 | 7.3 | Qwen3.8-Max, Doubao, Kimi, Pangu |
Three observations:
The middle is the most crowded and not the highest-scoring. ¥4–10 has 12 models averaging 84.2, because the band contains both GLM-5.3 (92.4) and scenario models such as Congrong (76.1) and Fuxi (77.5). The more crowded the band, the less you can select by band; you must read dimensions.
¥10–15 has only three models and the highest average (92.2). This is the sub-flagship sweet spot: Qwen3.5 94.6, GLM-5 92.0, SenseNova 89.9. Small sample, but none is weak.
¥15–21 averages lower than ¥10–15. Pangu 5.5 at ¥20 ranks #21 (81.6) and drags the band down. The specimen of expensive-but-not-strong: Pangu’s price includes industry authorisation and localisation, not tokens (see the industry exclusivity map). Buy it for general capability and much of the ¥20 pays for something you do not need.
Four pricing strategies
Classify the 26 by price behaviour:
1. Penetration pricing (¥0.5–2.5). MiniCPM ¥0.5, DeepSeek-V4.1-Flash ¥0.6, OCR 2 ¥0.8, InternVL ¥1.0, Qwen3.8-Flash ¥1.2, Keye ¥2, LongCat ¥2.5. The goal is volume or ecosystem share; this is also where each yuan buys the most points.
2. Mid-band pricing (¥4–10). Omni ¥4, MiMo ¥6, GLM-5.3 ¥9, Step ¥5, ERNIE ¥8, Hunyuan ¥8.6, MiniMax ¥8, Spark ¥6, Ant ¥5, JD ¥6, Fuxi ¥4, Congrong ¥7. The most crowded; differentiation comes from modality or scenario, not price.
3. Flagship pricing (¥12–20). Qwen3.5 ¥12, SenseNova ¥12, Doubao ¥15, Kimi ¥16, Qwen3.8-Max ¥18, Pangu ¥20. You buy capability ceiling, #1 position, long context or industry authorisation.
4. Zero/project pricing (¥0). Six models; see zero-price models. They do not compete on token price.
Procurement: set the band, then pick the frontier
The practical value of price cliffs is splitting selection into two steps:
Step one: set the budget band. Not “cheaper is better”, but “what can this scenario afford per month”. Once the band is set, cross-band comparison is meaningless — a ¥6 model and an ¥18 model are not the same product.
Step two: pick the frontier inside the band. In each band, only one or two models are “highest score at this price”; the rest sell modality, licence or industry solutions. If you only need score, pick the frontier; if you need modality/licence/scenario, pick the specialised component outside it.
Step three: compute total cost. Token price is one TCO line; image tokens, cache, volume discounts, on-prem and compliance all count. See cost per 1,000 images and the hidden bills of zero-price models.
A concrete suggestion: for “general + budget-sensitive”, ¥6–12 is the current sweet range — MiMo ¥6 (90.1), GLM-5.3 ¥9 (92.4), Qwen3.5 ¥12 (94.6); every ¥3 buys a real jump. For “flagship + #1”, the ¥15–18 premium buys 0.6–0.8 points and brand certainty, not an order-of-magnitude difference.
Boundaries
- Price is output token list price; it excludes input, image tokens, cache, volume discounts and on-prem. A true TCO changes the frontier.
- The 26 priced models are a snapshot. Vendor price changes move the cliffs; numbers are as of 2026-10-18.
- The frontier is two-dimensional (price + composite). Include eight dimensions, licence and context and it becomes nine-dimensional; that version is in cost per 1,000 images.
- Correlation is not causation. r = 0.44 only says “more expensive models average somewhat stronger”, not “paying more makes a model stronger”.
Reproduction
Frontier algorithm: sort by priceOut ascending, keep points whose composite is higher than every cheaper model; eight survive. Cliffs are the price and score differences between adjacent frontier points; bands are interval groups. Data from src/data/models.ts; scripts ship with the data; rerun after every expansion, and the price table updates in the changelog.
If this essay leaves one sentence: Chinese multimodal pricing is not a curve but a jump between anchors. Find your budget band first, then ask who is on the frontier inside it; cross-band comparison is comparing two different products.
FAQ about this essay
Can price proxy for capability?
Only weakly. Across 26 priced models r = 0.44, R² = 0.19; Pangu 5.5 at ¥20 ranks #21, and 20 cheaper models score higher.
Why set a budget band before comparing models?
Because cross-band comparison compares two products: ¥6 buys general or specialist capability, ¥18 buys rank, modality or compliance. Set the band, then find the frontier inside it; comparing score per yuan across bands underrates what flagships uniquely provide.
