Three things to remember
- Four models, four strategies in one window: Qwen3.8-Max raises price 50% for +0.8 points; GLM-5.3 cuts price 10% and rebalances capability (agents +4, generation −4); DeepSeek-V4.1-Flash cuts price 25% and turns a specialist into a generalist (agents +13, ecosystem +19); Qwen3.8-Flash delivers 92.5% of its predecessor’s score at one-tenth the price.
- All four arrive at evidence grade B. That is not our caution — it is a structural fact of Q4 2026: release windows compress into earnings season and third-party reproduction cannot keep up.
- GUI agents went from flagship feature to launch-window standard: three of the four declare a GUI modality. Q4 2026 is the adoption window for GUI agents — and the first window in which recovery rate matters more than success rate.
Eleven days, four models, one compressed release window
Between 30 September and 9 October 2026, DeepSeek, Zhipu and Alibaba released four models in eleven days: DeepSeek-V4.1-Flash, GLM-5.3, Qwen3.8-Max and Qwen3.8-Flash. That is no coincidence — a release window ahead of earnings season, training-cluster scheduling, and the shared instinct of “entering before evaluation bodies freeze annual data” compressed the window.
This essay paraphrases no launch post. Launch posts answer “how strong is this model”; buyers actually ask four other questions:
- Within one window, who raised prices and who cut them?
- Against each predecessor, which dimension gave capability to which?
- With all four at evidence grade B, what should someone about to sign a contract do?
- What structural move in Chinese multimodal did this window expose?
Four questions, four tables. Every number is recomputable from the 256 score points of the 32 models on this site; the script ships with the data.
The same-window alignment table
| DeepSeek-V4.1-Flash | GLM-5.3 | Qwen3.8-Flash | Qwen3.8-Max | |
|---|---|---|---|---|
| Release date | 2026-09-30 | 2026-10-02 | 2026-10-09 | 2026-10-09 |
| Vendor | DeepSeek | Zhipu AI | Alibaba | Alibaba |
| Output price | ¥0.6 | ¥9 | ¥1.2 | ¥18 |
| Composite | 87.9 | 92.4 | 87.5 | 95.4 |
| Overall rank | #12 | #4 | #13 | #1 |
| Evidence grade | B | B | B | B |
| Licence | MIT open weights | Closed + on-prem licence | Apache-2.0 open weights | Closed commercial API |
| Context | 128K | 400K | 256K | 1M (256K native) |
| Modalities | Text/image/video | Text/image/video/GUI | Text/image/video/GUI | Text/image/video/audio/GUI |
| Predecessor | DeepSeek-OCR 2 | GLM-5 | Qwen3.5-397B | Qwen3.5-397B |
The four span 30× in price (¥0.6 to ¥18) and 7.9 points in composite. Inside one release window, the four vendors answered “who is this model sold to” in completely different ways — these are not four rivals but four procurement identities.
The lineage delta table: the true shape of an upgrade
“Newer is stronger” is the laziest conclusion. The delta table breaks each upgrade into per-dimension differences, all computed from our scores. It answers: what did this upgrade take, and what did it give.
GLM-5.3 versus GLM-5: a product rebalance
| Dimension | GLM-5 | GLM-5.3 | Change |
|---|---|---|---|
| Agent | 91 | 95 | +4 |
| Ecosystem | 92 | 97 | +5 |
| Video | 88 | 88 | 0 |
| Perception | 93 | 93 | 0 |
| Chinese | 95 | 95 | 0 |
| Document | 95 | 94 | −1 |
| Chart | 92 | 91 | −1 |
| Generation | 89 | 85 | −4 |
| Composite | 92.0 | 92.4 | +0.4 |
| Price | ¥10 | ¥9 | −10% |
A minor version number with major movement: agents +4, ecosystem +5, generation −4. This is a deliberate reallocation of budget from generation toward agents and delivery — the −4 on generation is the largest voluntary concession in any same-family generational step here. The reading is direct: Zhipu believes enterprise buyers in late 2026 are paying for tool calls and private deployment, not image generation. If you buy GLM-5.3 for image generation, you bought the tier it deliberately gave up.
Qwen3.8-Max versus Qwen3.5-397B: gaps closed, and one taken
| Dimension | Qwen3.5-397B | Qwen3.8-Max | Change |
|---|---|---|---|
| Agent | 92 | 95 | +3 |
| Perception | 96 | 97 | +1 |
| Chart | 95 | 96 | +1 |
| Video | 92 | 93 | +1 |
| Chinese | 94 | 95 | +1 |
| Document | 97 | 97 | 0 |
| Generation | 90 | 90 | 0 |
| Ecosystem | 98 | 96 | −2 |
| Composite | 94.6 | 95.4 | +0.8 |
| Price | ¥12 | ¥18 | +50% |
What it fixed is the weak side (agents +3, video +1); what it gave is ecosystem (−2: Qwen3.5-era deployment-tooling advantages reallocated to the open smaller sizes). The price is +50% — the most expensive generational price increase on this board — and it bought +0.8 of composite. Whether that exchange rate pays depends on whether your workload sits exactly on agents and video.
DeepSeek-V4.1-Flash versus DeepSeek-OCR 2: from specialist to generalist
| Dimension | DeepSeek-OCR 2 | DeepSeek-V4.1-Flash | Change |
|---|---|---|---|
| Ecosystem | 76 | 95 | +19 |
| Agent | 72 | 85 | +13 |
| Generation | 60 | 72 | +12 |
| Video | 70 | 78 | +8 |
| Perception | 84 | 90 | +6 |
| Chart | 88 | 92 | +4 |
| Document | 99 | 95 | −4 |
| Chinese | 90 | 86 | −4 |
| Composite | 81.9 | 87.9 | +6.0 |
| Price | ¥0.8 | ¥0.6 | −25% |
The most violent change of the four, and the biggest repositioning: DeepSeek turned a “layout-parsing specialist” into a “general native multimodal model” — document −4, ecosystem +19, agents +13. It gave up 4 points of documents and bought catch-up on the other seven dimensions. For buyers: first-pass document parsing should still use OCR 2 (the 99 stands), but second-pass review can now use its sibling V4.1-Flash instead of a flagship — that combination costs ¥0.72 + ¥0.54 per thousand images, against ¥0.72 + ¥16.29 for “flagship review”.
Qwen3.8-Flash versus Qwen3.5-397B: one-tenth the price, 92.5% of the score
No dimension-by-dimension comparison here — these are different products. One ratio suffices: Qwen3.8-Flash takes one-tenth of Qwen3.5-397B’s price (¥1.2 vs ¥12) for 92.5% of its composite (87.5 vs 94.6), and is one of the two most permissively licensed of the four new models (Apache-2.0). It is the first time a Flash-class model reaches tier A on this board, which redrew the price-performance Pareto frontier.
The grade-B wave: four new models, one shared uncertainty
Four models, four Bs. Same reason across the board: under 30 days old, no third-party reproduction, no 30-day stability record. That is not our caution — release windows compress into earnings season and third-party evaluation cannot keep vendor pace — a structural fact of Q4 2026.
Quantifying the B helps. Under our rules, a grade-B score carries a reasonable range of about ±1.5 points:
| Model | Published | Reasonable range | Meaning |
|---|---|---|---|
| Qwen3.8-Max | 95.4 | 93.9–96.9 | May swap with Doubao (95.2) or open a gap |
| GLM-5.3 | 92.4 | 90.9–93.9 | May touch the lower edge of S+ or fall back to A+ |
| DeepSeek-V4.1-Flash | 87.9 | 86.4–89.4 | May enter the upper edge of A− |
| Qwen3.8-Flash | 87.5 | 86.0–89.0 | Same |
The operating advice has one dividing line: project cycles longer than one evaluation cycle (about three months) → wait for reproduction; cycles that cannot wait → buy back certainty with the grade-A runner-up. The runners-up are ready-made: Doubao Seed 2.0 Pro (95.2 / A) and GLM-5 (92.0 / A). Trading 0.2 points for an A is usually a good deal in a three-year contract — that is the entire logic of gate five in the veto-gate tier list.
GUI agents: the window’s most underestimated move
Of the four new models, three declare a GUI-control modality (GLM-5.3, Qwen3.8-Flash, Qwen3.8-Max); only DeepSeek-V4.1-Flash does not. Before this window, exactly one model on the whole board declared GUI: Kimi K3.
So GUI agents went from a flagship’s single feature to a launch-window standard declaration in eleven days. That matters more than any single score, for the reason written up in the GUI agent benchmark: public benchmarks stop discriminating at 95, and the real watershed is failure recovery — GLM-5.3 78% > Qwen3.8-Max 71% > Kimi K3 64%. When GUI declarers go from one to four, “can tap buttons” stops being a selling point next year; “recovers when it errs” becomes one. Writing recovery rate into your acceptance criteria is worth far more than writing “supports GUI” into your requirements doc.
Eight procurement identities, eight bills
Put the four frontier models and the same-cycle industry entries back into their positions and this window is really a complete procurement map:
| If your business is | Pick | Per 1,000 images (1024px) | Why |
|---|---|---|---|
| High-volume document pipelines | DeepSeek-V4.1-Flash | ¥0.54 | Document 95 + MIT licence; the cost baseline |
| High-frequency image tagging / moderation | Qwen3.8-Flash | ¥1.09 | Ecosystem 97 + Apache-2.0; fastest delivery |
| Domestic-chip / on-prem | GLM-5.3 | ¥8.14 | Full Ascend stack + on-prem licence; unmatched |
| Capability ceiling first | Qwen3.8-Max | ¥16.29 | Composite leader; chart/document both 96+ |
| Financial instruments / compliance | Ant Ling 2 | ¥4.53 | Chinese 92 plus finance scenario corpora |
| E-commerce images / supply chain | JD Yanxi 2 | ¥5.43 | Exclusive product-image and SKU corpora |
| Games / short-video content | NetEase Fuxi 2 | ¥3.62 | Generation 82 + video 84; content-oriented |
| Surveillance | Hikvision Guanlan 3 | Industry licensing | The only surveillance-video option; ships with hardware |
Note the 8× price gap between rows three and four — “domestic substitution” and “capability ceiling” are still two markets in October 2026, not two price points. Note also the difference in kind between the last four rows and the first four: the first four buy “general capability”, the last four buy “scenario corpora” — the latter are almost all dominated on the Pareto frontier, yet each is a “no second place” on the industry exclusivity map. That is the hardest structural fact this release window produced, and the reason G2 is the second gate in the enterprise selection playbook.
What to revisit in three months
Three pre-registered observation points, to stop this analysis decaying into another launch-post paraphrase:
- Does third-party reproduction converge? Do all four Bs rise to A before January 2027, and do the scores stay within ±1.5? Any escape means systematic bias between vendor figures and independent measurement.
- Does the frontier anchor move? The current value anchor is DeepSeek-V4.1-Flash (¥0.54/87.9). If a next-gen Flash compresses the same capability below ¥0.3, the frontier shape is rewritten — at which point this essay’s tables expire and the new ones appear in cost per 1,000 images.
- Does GUI recovery rate become a public benchmark? If AndroidWorld or OSWorld adds a “failure recovery” metric next version, this essay’s exclusive artefact becomes public property — which is exactly what this site is for: turning exclusives into common knowledge, then going to find the next exclusive.
Boundaries and method
The delta table’s scores are our editorial mappings, not raw figures from any public benchmark; predecessor and successor are scored under identical weights for comparability. The ±1.5 range for grade B is our empirical estimate for the “vendor-reported plus partial reproduction” evidence class, not a statistical confidence interval. All numbers are revised through the changelog, and revisions are traceable.
FAQ about this essay
All four are grade B — can I pick one now?
It depends on your project cycle. For anything longer than a quarter, waiting for third-party reproduction usually pays: a grade-B score can drift between 91 and 93.5, while for a grade-A model we are comparatively certain. If the cycle cannot wait, take the grade-A runner-up (Doubao 95.2, GLM-5 92.0) and buy back certainty — leave version churn to the next contract cycle.
What is the lineage delta table and how do I use it?
It is a per-dimension delta table comparing each new model against its same-family predecessor, computed directly from our 208 score points. Use it to read the true shape of an upgrade: GLM-5.3 versus GLM-5 is a product rebalance — agents +4, ecosystem +5, generation −4 — not a uniform gain; Qwen3.8-Max versus Qwen3.5-397B shows ecosystem −2, meaning the generation took as well as gave. Read the delta table before upgrading to avoid the default assumption that newer is uniformly better.
