The price-performance Pareto frontier: which of the 22 priced models are strictly dominated

Unit price is not value, and most boards publish only unit price

“Price per million tokens” is the number vendors love most and the one least useful in selection. It does not tell you how many tokens an image eats, how many times dearer output is than input, or whether the capability at that price is worth it.

We change the unit: how much does it cost to process 1,000 images. Then we place it on the same plane as the composite from the 2026 China native multimodal LLM tier list and draw a Pareto frontier. That frontier tells you something far more important than “who is cheapest”: which models should never be selected at this price point.

Calculation basis (identical to the site’s cost calculator)

All numbers come from one code path; change the resolution in the calculator and you reproduce them:

• Input cost = 1,000 × (image tokens + 120) × (output price × 25%) ÷ 1,000,000
• Output cost = 1,000 × 600 × output price ÷ 1,000,000
• Each image carries 120 tokens of prompt; input priced at 25% of output
• Cache hits, retries and volume discounts excluded
• Models without public prices are excluded (4 of 26 here: Emu3.5, BAGEL-2, WITA-Omni, NeuroXess-N1)

Image tokens use four resolution presets: 720p short side 340, 1024px 1100, 1440px 2500, 4K/long 4200.

The 1024px tier: the real bill for 1,000 images

Model	Output price	Cost per 1,000	Composite	Overall
MiniCPM-V 4.5	¥0.5	¥0.45	79.6	#24
DeepSeek-V4.1-Flash	¥0.6	¥0.54	87.9	#12
DeepSeek-OCR 2	¥0.8	¥0.72	81.9	#20
InternVL 4.0	¥1.0	¥0.91	85.0	#16
Qwen3.8-Flash	¥1.2	¥1.09	87.5	#13
Keye-VL-2.0	¥2	¥1.81	86.0	#14
LongCat-Next	¥2.5	¥2.26	89.2	#9
Qwen3-Omni-30B	¥4	¥3.62	88.9	#10
Step-3	¥5	¥4.53	84.7	#17
MiMo-V2.5	¥6	¥5.43	90.1	#7
Spark X2	¥6	¥5.43	84.5	#18
ERNIE 5.1	¥8	¥7.24	88.1	#11
MiniMax M2.5	¥8	¥7.24	83.8	#19
Hunyuan-Vision	¥8.6	¥7.78	85.2	#15
GLM-5.3	¥9	¥8.14	92.4	#4
GLM-5	¥10	¥9.05	92.0	#6
Qwen3.5-397B	¥12	¥10.86	94.6	#3
SenseNova 6.5 Pro	¥12	¥10.86	89.9	#8
Doubao Seed 2.0 Pro	¥15	¥13.57	95.2	#2
Kimi K3	¥16	¥14.48	92.3	#5
Qwen3.8-Max	¥18	¥16.29	95.4	#1
Huawei Pangu 5.5	¥20	¥18.10	81.6	#21
Ant Ling 2	¥5	¥4.53	80.8	#22
MiniCPM-V 4.5	¥0.5	¥0.45	79.6	#26
JD Yanxi 2	¥6	¥5.43	78.0	#28
NetEase Fuxi 2	¥4	¥3.62	77.5	#29
Cloudwalk Congrong 2	¥7	¥6.33	76.1	#32

Cheapest to dearest is a 40× spread (¥0.45 vs ¥18.10). The composite spread is only 19.3 points (76.1 vs 95.4). That ratio is the starting point of this whole essay: price dispersion is far larger than capability dispersion. So pricing contains a great deal that has nothing to do with capability — brand, channel, industry licensing, compute self-sufficiency.

The newly listed InternVL 4.0 lands at ¥0.91/85.0: 26% dearer than DeepSeek-OCR 2 for +3.1 composite, and 17% cheaper than Qwen3.8-Flash for −2.5. It does not reach the frontier, but it is the first time the open-source camp has brought perception 96 under ¥1 — meaningful for an open-source stack, irrelevant to the frontier.

The four industry models’ price points are themselves informative: Ling 2 ¥4.53 (80.8), Yanxi 2 ¥5.43 (78.0), Fuxi 2 ¥3.62 (77.5), Congrong 2 ¥6.33 (76.1) — all four land below-right of existing points, and none is simultaneously cheaper than some frontier point and stronger than it. Industry pricing contains scenario delivery, not token unit price, which keeps them off this frontier by construction.

The Pareto frontier: only 8 of 22 should not be swapped out

Treat cost as minimise and composite as maximise, then run a two-dimensional Pareto filter. A model is dominated if another exists that is cheaper (or no dearer) and stronger (or no weaker), strictly better on at least one. Such a model has no reason to be selected — however good its marketing page reads.

14 of the 22 priced models are dominated. Eight remain on the frontier:

Frontier slot	Model	Per 1,000	Composite	What the money buys
1	MiniCPM-V 4.5	¥0.45	79.6	Extreme low price; capability traded for cost
2	DeepSeek-V4.1-Flash	¥0.54	87.9	The field’s biggest jump: +¥0.09 buys +8.3 points
3	LongCat-Next	¥2.26	89.2	Generation 92, second on the site
4	MiMo-V2.5	¥5.43	90.1	Video 93; the cheapest strong video hand on the frontier
5	GLM-5.3	¥8.14	92.4	Full Ascend stack + agent 95
6	Qwen3.5-397B	¥10.86	94.6	All eight dimensions top-three
7	Doubao Seed 2.0 Pro	¥13.57	95.2	Perception 97, video 96, Chinese 97
8	Qwen3.8-Max	¥16.29	95.4	The composite leader

The shape of this frontier speaks for itself: the first two slots are ¥0.09 apart but 8.3 points apart — DeepSeek-V4.1-Flash reaches 87.9 at a ¥0.6 output price, the steepest segment on the entire frontier. In reverse, from GLM-5.3 (¥8.14/92.4) to Qwen3.8-Max (¥16.29/95.4) the price doubles for 3.0 points. Marginal returns decay cleanly.

If your budget is capped at ¥1 per thousand, the frontier’s answer is unique: DeepSeek-V4.1-Flash. Not “one of the best-value options” — there is no second frontier point in that budget band.

A finding across three expansions: the frontier is immune to new arrivals

This is the most stable by-product of the 24 → 26 → 32 expansions: the frontier’s eight members did not change at all.

• InternVL 4.0 (¥0.91/85.0) is two-dimensionally dominated by DeepSeek-V4.1-Flash (¥0.54/87.9) — cheaper and stronger, so no frontier slot.
• The four industry models all land below-right of existing points: Ling 2 (¥4.53/80.8) is suppressed by both DeepSeek-V4.1-Flash at ¥0.6 and Qwen3-Omni-30B at ¥4; Yanxi 2, Fuxi 2 and Congrong 2 are dominated by MiMo-V2.5 and others.
• NeuroXess-N1, Guanlan and Zidong Taichu have no public price and do not enter under our rules.

That is not luck. To join this frontier, a new model must be simultaneously “cheaper than some frontier point” and “stronger than it”. Two Flash models dragged the price line to about ¥0.6 in late 2026, so every arrival since either beats that price band or lands below-right of existing points. The frontier’s stability is itself the conclusion: the ¥0.6–1.2 price band is saturated as of October 2026, and the next frontier point requires the next price cut, not the next capability bump.

But a 2D frontier is too loose: switch to 9-dimension strict dominance

The analysis above has a problem — it compresses eight capability dimensions into one composite, and the composite is our weighting. In other words, the 2D Pareto conclusion depends on the weight assumption, which is exactly what we tried to dismantle in the 2,000-run weight perturbation.

So we run it a second way: no composite, strict dominance over price plus all eight dimensions. A model is 9-dimension strictly dominated if another exists that is no dearer, no weaker on any of the eight dimensions, and strictly better on at least one.

This standard is far harsher, and the result is far cleaner: only 8 of 26 are strictly dominated.

Dominated	Price / composite	Dominated by
SenseNova 6.5 Pro	¥12 / 89.9	Qwen3.5-397B (same ¥12, no dimension below)
Hunyuan-Vision	¥8.6 / 85.2	MiMo-V2.5 (¥6), Qwen3-Omni-30B (¥4), ERNIE 5.1 (¥8)
Spark X2	¥6 / 84.5	MiMo-V2.5 (same ¥6), Qwen3-Omni-30B (¥4)
MiniMax M2.5	¥8 / 83.8	MiMo-V2.5 (¥6, no dimension below)
Huawei Pangu 5.5	¥20 / 81.6	13 models at once; the cheapest is Qwen3.8-Flash (¥1.2)
JD Yanxi 2	¥6 / 78.0	MiMo-V2.5 (same ¥6), Qwen3-Omni-30B (¥4), InternVL 4.0 (¥1)
Cloudwalk Congrong 2	¥7 / 76.1	MiMo-V2.5 (¥6), LongCat-Next (¥2.5)
NetEase Fuxi 2	¥4 / 77.5	LongCat-Next (¥2.5), Qwen3-Omni-30B (same ¥4)

MiMo-V2.5 is the hidden winner of this table. 7th overall, evidence grade B, discussed by almost nobody — yet under 9-dimension strict dominance it beats five opponents at once (Hunyuan, Spark X2, MiniMax M2.5, Yanxi 2, Congrong 2), at an equal or lower price. It is on no one’s “top models” list, yet it is on the “no reason not to pick it” list.

Huawei Pangu 5.5 is the other extreme. ¥20 is the highest unit price here, 11% above the leader Qwen3.8-Max, while its composite is 81.6, 21st, strictly dominated by 13 models. That is not a capability problem — it has real industry deliveries (typhoon-track error −30%, 72-hour fault warning on mine equipment). It is the specimen of pricing decoupled from general capability: its price contains industry licensing and localised service, not tokens. If that is what you are buying, Pareto analysis does not apply to you; if you thought you were buying general multimodal capability, this table is stopping you.

From 18 to 8: how the ten escapees survive

The 2D frontier marks 18 dominated; 9-dimension strict dominance only 8. The ten that escape survive on one irreplaceable dimension:

Model	Dominated in 2D	What saves it
Kimi K3	Doubao Seed 2.0 Pro (cheaper, stronger)	Agent 95, tied with the leader — a tie, not a lead
GLM-5	Qwen3.5-397B, Doubao	document 95, chart 92, generation 89 — all above every cheaper rival
Qwen3-Omni-30B	LongCat-Next, MiMo-V2.5	video 90, chinese 91, ecosystem 90
ERNIE 5.1	MiMo-V2.5, Qwen3-Omni-30B	chinese 93
Keye-VL-2.0	Qwen3-Omni-30B, MiMo-V2.5	video 94, 12 points above the strongest same-price rival
DeepSeek-OCR 2	DeepSeek-V4.1-Flash	document 99, first on the site, 4 points clear
Step-3	MiMo-V2.5, Qwen3-Omni-30B	No strictly leading dimension — survives on ties alone
Qwen3.8-Flash	DeepSeek-V4.1-Flash	video 82, agent 87, generation 80, chinese 88, ecosystem 97 — five strict leads
InternVL 4.0	DeepSeek-V4.1-Flash	perception 96, chinese 89 — two strict leads over every cheaper rival
Ant Ling 2	DeepSeek-V4.1-Flash, Qwen3-Omni-30B	chinese 92 — third on the site, above every cheaper rival

This table rewards row-by-row reading.

DeepSeek-OCR 2’s reprieve is the most typical. 20th overall, composite 81.9, crushed in 2D by its sibling DeepSeek-V4.1-Flash (12th, 87.9, slightly dearer). But its document score is 99, first on the site — two points above the joint second Qwen3.5-397B and Qwen3.8-Max (both 97) and four above DeepSeek-V4.1-Flash’s 95. If your business is Chinese document extraction, those four points are its entire reason to exist — at ¥0.8 per thousand. It is not “the cheap substitute”; it is the optimum on that single dimension.

Qwen3.8-Flash’s reprieve is the most counter-intuitive. At ¥1.2 it is double DeepSeek-V4.1-Flash’s ¥0.6 with a composite 0.4 lower. In 2D it is fully dominated and looks like it has no reason to exist. But across eight dimensions it strictly leads on five: video 82 vs 78, agent 87 vs 85, generation 80 vs 72, chinese 88 vs 86, ecosystem 97 vs 95.

That last 97 is the point. The ecosystem dimension measures SDK completeness, deployment tooling, community size, documentation quality and third-party integrations — it carries only 8% weight in the composite, yet it decides how many person-weeks your engineering team spends. DeepSeek-V4.1-Flash is half the price but you build the inference stack yourself; Qwen3.8-Flash is double the price and works out of the box on Bailian, with day-one vLLM/SGLang support and Apache-2.0 with no strings. The extra ¥0.6 per thousand buys delivery speed, not intelligence. Whether that trade is worth it depends on your launch date, not on the tier list.

InternVL 4.0’s reprieve is the newest. It is 2D-dominated by DeepSeek-V4.1-Flash (69% dearer, composite 2.9 lower), yet on perception it is 96 against 90 — a 6-point gap, and the third-highest perception score on the site. For an open-source stack hunting a perception component under ¥1, it is the only option; closed-source shops never look at it. The model’s very existence is a reminder that the frontier graph is biased for open-source users, because the frontier sorts by composite, while an open-source buyer’s utility function puts licence and self-hosting at close to 40% — and neither lives in the eight dimensions.

Kimi K3 and Step-3 survive most fragilely. Neither strictly leads every cheaper rival on any dimension; they only tie. Kimi K3’s agent 95 ties Qwen3.8-Max and GLM-5.3 — it survives on being “as strong as the strongest but ¥2 cheaper”. Step-3 is weaker still, pure tie. Ties are survival reasons that the next version update erases.

The three new industry models did not escape the 9-dimension test, and that is itself the conclusion. Yanxi 2 (¥6/78.0) is strictly dominated by same-priced MiMo-V2.5 — leading on none of eight dimensions; Fuxi 2 (¥4/77.5) by cheaper LongCat-Next and same-priced Qwen3-Omni-30B; Congrong 2 (¥7/76.1) by MiMo-V2.5 and LongCat-Next. All three survive entirely on the ninth dimension and beyond — scenario corpora, industry channels, hardware synergy — and none of those lives in the eight capability dimensions. The 9-dimension test answers one question: “as a general-capability purchase, is there any reason to select you?”. The right answer for industry models was never “to be selected” but “to have no substitute in a specific scenario” — and that question needs a different table, the industry exclusivity map.

Ant Ling 2 is the only new model that escaped the 9-dimension test. It is 2D-dominated by DeepSeek-V4.1-Flash (¥0.6) and Qwen3-Omni-30B (¥4), but its Chinese 92 sits above every cheaper rival (the strongest, ERNIE 5.1 at 93, is 60% dearer). Translated: when you are buying “Chinese capability” as a specific product, the frontier graph fails — the frontier sorts by composite, and Chinese capability carries only 6% of the composite.**

Does resolution reshuffle the frontier?

It changes absolute values, not membership. Running all four resolutions:

Resolution	Cheapest	Dearest	Tail ratio	Frontier members
720p (340 tokens)	MiniCPM-V 4.5 ¥0.36	Pangu 5.5 ¥14.30	39.7×	Same 8
1024px (1100)	MiniCPM-V 4.5 ¥0.45	Pangu 5.5 ¥18.10	40.2×	Same 8
1440px (2500)	MiniCPM-V 4.5 ¥0.63	Pangu 5.5 ¥25.10	39.8×	Same 8
4K/long (4200)	MiniCPM-V 4.5 ¥0.84	Pangu 5.5 ¥33.60	40.0×	Same 8

The eight members are identical across all four resolutions, with absolute amounts scaled up uniformly, because every model’s image-token pricing is isomorphic — resolution multiplies everyone equally. That means: resolution is not a variable you need to re-run selection for. Do the Pareto analysis at 1024px, then move your images to 4K and the frontier order holds.

But the absolute values move, hard: at 4K, Qwen3.8-Max costs ¥30.24 per thousand images, 2.35× the 720p figure. If your business is long images, high-resolution receipts or engineering drawings, cut resolution before you switch models — compressing 1440px to 1024px saves about 33%, the same as switching from Qwen3.8-Max to Qwen3.5-397B (about 33%), and the former costs you no flagship capability.

Three conclusions you can copy

One, under ¥1 per thousand: DeepSeek-V4.1-Flash, no second option. Unless your core scenario is pure document extraction, where DeepSeek-OCR 2 (¥0.8) is stronger on the single dimension; unless delivery speed and ecosystem matter, where Qwen3.8-Flash (¥1.09) earns its 2× premium; unless your stack is open source and you need a permissive licence, where InternVL 4.0 (¥0.91) is the open camp’s only strong off-frontier option.

Two, ¥8–11 per thousand: the most contested and most error-prone band. GLM-5.3 (¥8.14 / 92.4) and GLM-5 (¥9.05 / 92.0) are both on the frontier, ¥0.91 apart for 0.4 points. The real decision variable is not the score but GLM-5.3’s full Ascend stack — for domestic-chip substitution it has no rival in this band.

Three, at any budget, read the 9-dimension strict dominance table first. If your candidate is among those five, swap it out and skip the POC. SenseNova 6.5 Pro and Qwen3.5-397B at the same ¥12 with SenseNova behind on all eight dimensions — that is not “different strengths”, it is a pure loss.

Every number here is computed by analyze3.mjs from raw scores and prices, sharing one formula with the site’s cost calculator. Change resolution or output tokens and it reproduces instantly.

Evidence grade is deliberately absent from this frontier graph. Cheap and strong are capability questions; reproducibility is a certainty question — mixing them muddles both. The separated rubric is on the methodology page.