2026 CHINA AI
PRICE_
WAR_
REPRICE.
In a five-day window, three of China's top AI labs made moves that look contradictory on the surface. DeepSeek raised API prices by as much as 1,100% on certain tiers. Alibaba, on the same week, open-weighted a 2.4-trillion-parameter flagship model it had never released before. Zhipu AI shipped GLM-5.3, boosting coding benchmarks by roughly 6x using the exact same base model as its predecessor — no retraining involved. Pain: headline percentages do not map to a real bill, and "Chinese model = cheapest" no longer holds. Takeaway: the labs are shifting from competing on price alone to competing on pricing power. Structure: timeline, price and spec tables, three strategies, head-to-head rates, unverified claims, a five-step checklist, and FAQ.
1. Three constraints that actually changed this week
- The 1,100% headline is one billing line, not the whole invoice. It applies to peak-hour cache-hit input — the tier that started nearest to free. Output, which dominates most real bills, rose 350%. Cache-miss input rose 200%.
- The official API is no longer automatically the cheapest host. At peak hours, DeepSeek's own list price now sits above several third-party resellers (GMI Cloud, Novita, and others currently list V4 Pro below the new official peak). Off-peak V4-Pro is still well below Claude Opus 5, but Qwen3.8-Max international ($2/$6) and OpenAI GPT-5.6 Luna ($0.20/$1.20) undercut DeepSeek's off-peak rate on at least one dimension.
- Open weights are not Apache 2.0. Alibaba attached a custom Qwen3.8-Max License: any Model-as-a-Service or AI Work Assistant business earning over $50 million in any 12-month period must negotiate a separate commercial license. Claims that the license bans US, EU, UK, or South Korea downloads are false — the published text has no geographic clause.
2. Timeline: what happened, and when
| Date | Event |
|---|---|
| Jul 16, 2026 | Moonshot AI open-weights Kimi K3 (2.8T parameters), drawing US security scrutiny |
| Aug 2–3, 2026 | Alibaba previews, then launches, Qwen3.8-Max as a hosted API |
| Aug 10, 2026 | Meta releases Muse Glimmer (30B, Apache 2.0), teases open weights for flagship Muse Spark 1.2 |
| Aug 12, 2026 | Alibaba publishes Qwen3.8-2.4T-A95B open weights on Hugging Face/ModelScope; xAI ships Grok 4.6 |
| Aug 13, 2026 | DeepSeek-V4-Pro goes GA and announces a price increase effective Aug 17; Google ships discounted Gemini 3.7 Flash |
| Aug 14, 2026 | Zhipu ships GLM-5.3, reusing GLM-5.2's 743B base |
| Aug 17, 2026, 00:00 Beijing time | DeepSeek's new pricing takes effect |
Zoom out further: on Jul 30, OpenAI cut prices on its cheapest tier (GPT-5.6 Luna, down 80%), then on Aug 6–7 made Luna the free default with unlimited text chats. While Chinese labs were raising prices and opening flagship weights, US labs were cutting prices and going free at the consumer layer — at the exact same time. That is two sides of the same pricing fight.
3. The numbers: what actually changed
DeepSeek's price hike (effective Aug 17, 00:00 Beijing time; peak hours 9am–12pm and 2pm–6pm Beijing time)
| Billing item (per 1M tokens) | Old price | New off-peak | New peak | Peak increase |
|---|---|---|---|---|
| V4-Flash cache hit (input) | ¥0.02 | ¥0.05 | ¥0.10 | ~400% |
| V4-Flash cache miss (input) | ¥1.0 | ¥1.5 | ¥3.0 | 200% |
| V4-Flash output | ¥2.0 | ¥4.5 | ¥9.0 | 350% |
| V4-Pro cache hit (input) | ¥0.025 | ¥0.15 | ¥0.30 | ~1,100% |
| V4-Pro cache miss (input) | ¥3.0 | ¥4.5 | ¥9.0 | 200% |
| V4-Pro output | ¥6.0 | ¥13.5 | ¥27.0 | 350% |
The headline 1,100% figure applies specifically to peak-hour cache-hit input. Independent cost modeling found that a realistic heavy-usage workload (roughly 84M tokens/month, mostly off-peak, half cache hits) sees a bill increase closer to 1.8x — real, but far below the scariest headlines.
Qwen3.8-2.4T-A95B (Qwen3.8-Max open weights): key specs
| Spec | Detail |
|---|---|
| Parameters | 2.4T total, 95B active per token (MoE, 512 experts, 10 routed + 1 shared) |
| Context window | 262,144 tokens native (open checkpoint), extendable to ~1.01M; hosted Max version defaults to 1M |
| Release cadence | Preview Aug 2 → API live Aug 3 → open weights Aug 12 |
| API pricing (international) | $2/M input, $6/M output |
| License | Not Apache 2.0 — a custom Qwen3.8-Max License |
| Why it matters | First time Alibaba has open-weighted a Max-tier flagship; Qwen3.5/3.6/3.7 Max stayed API-only |
GLM-5.3 vs GLM-5.2: same base model, post-training only
| Benchmark | GLM-5.2 | GLM-5.3 | Change |
|---|---|---|---|
| Terminal-Bench 3.0 | 4.6% | 28.3% | +23.7 pts |
| DeepSWE v1.1 | 46.2% | 66.9% | +20.7 pts |
| Agents' Last Exam (CLI) | 23.8% | 28.5% | +4.7 pts |
| CyberGym | 77.2% | 84.5% | +7.3 pts |
| AutomationBench | 26.2% | 48.2% | +22.0 pts |
These are Zhipu's own reported numbers — no independent third-party re-run has been published yet. GLM-5.3 still trails GPT-5.6 Sol (34.6%) and Claude Fable 5 (33.7%) on Terminal-Bench 3.0; it is a top open-weight result, not an outright frontier win.
4. Breaking down the three strategies
DeepSeek: time-of-day pricing is a capacity problem, not a strategy pivot
The easiest misread is "China's cheapest model finally caved to margin pressure." The structure reads more like a company making compute constraints visible in the price sheet for the first time. Flat, always-cheap pricing worked as a customer-acquisition tool as long as GPU capacity kept pace with demand. Once usage grew exponentially and capacity did not, something had to become explicit — and "encouraging more flexible workload scheduling" is corporate-speak for "peak-hour compute is now scarce, please shift your load yourself."
One detail international coverage mostly missed: at peak hours, DeepSeek's official API is now higher than several third-party resellers. The assumption that the official API is always the cheapest way to run DeepSeek has been broken for the first time.
Alibaba: open weights buy ecosystem goodwill; a custom license protects the revenue ceiling
Alibaba did two things simultaneously: it published the full 2.4T-parameter checkpoint for free download, and it attached a custom license that requires any Model-as-a-Service or AI Work Assistant business earning over $50 million in any 12-month period to negotiate a separate commercial license, and requires products with 100M+ monthly active users or $20M+ in monthly revenue to prominently display the model's name.
The logic: give away the weights to win developer mindshare (especially internationally), while keeping pricing leverage over the handful of companies capable of building a competing inference business on top of it. That is a materially different bet than Meta's Muse Glimmer, which ships under unrestricted Apache 2.0.
One rumor worth killing explicitly: claims that Alibaba's license bans downloads from the US, EU, UK, and South Korea are false. The published license text contains no geographic or territorial clause of any kind.
GLM-5.3: no new base model, just a bigger post-training bet
The most interesting fact is the method: same 743B-parameter base as GLM-5.2, no retraining, and a roughly 6x jump on Terminal-Bench 3.0 (4.6% → 28.3%) purely from scaling up reinforcement learning environments in post-training. As pretraining scaling laws show diminishing returns, post-training RL scale is becoming an independent performance lever with a much lower cost floor than retraining a new foundation model. That is a meaningfully lower barrier to entry for mid-tier labs.
5. Head-to-head: is DeepSeek still the cheapest frontier-class model?
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Open weights? |
|---|---|---|---|
| DeepSeek V4-Pro (peak) | ¥9.0 (~$1.26) | ¥27.0 (~$3.78) | No |
| DeepSeek V4-Pro (off-peak) | ¥4.5 (~$0.63) | ¥13.5 (~$1.89) | No |
| Qwen3.8-Max (international API) | $2.00 | $6.00 | Yes (custom license) |
| OpenAI GPT-5.6 Luna | $0.20 | $1.20 | No |
| Claude Opus 5 (implied, per Alibaba's own comparison ratio) | ~$5.00 | ~$25.00 | No |
RMB-to-USD conversion at ~¥7.15/$1, approximate. Even after the hike, DeepSeek V4-Pro's off-peak rate is still well below Claude Opus 5, but it is no longer the outright cheapest option — both Qwen3.8-Max's international pricing and OpenAI's Luna now undercut DeepSeek's off-peak rate. "Chinese model = cheapest model" was true for most of 2025 and early 2026; it is not a safe assumption anymore.
Quotable: V4-Pro peak cache-hit input ¥0.025 → ¥0.30 (~1,100%); output ¥6.0 → ¥27.0 (350%); Qwen3.8-Max international $2/$6; GLM-5.3 Terminal-Bench 3.0 4.6% → 28.3% on the same 743B base, no retraining.
6. What is disputed or unverified
- The 1,100% headline is technically accurate but misleading without context. It applies only to peak-hour cache-hit input. Output — the cost that dominates most real bills — rose 350%.
- Claims that Qwen3.8-Max runs on Alibaba's in-house Zhenwu M890 chips (and "Pangu AL128" supernodes), reported by several Chinese financial outlets, have not been independently confirmed by Alibaba's own technical documentation or third-party benchmarks. Treat as vendor-adjacent, unverified reporting.
- GLM-5.3's reported discovery of a "serious vulnerability" in Cursor comes from VentureBeat and Zhipu's own disclosure; specific technical details have not been made public. Read as vendor-sourced, not independently audited.
- Reports that China's Ministry of Commerce may be preparing retaliatory export controls on AI/semiconductor technology are speculative and sourced to unconfirmed media reports, not an official announcement.
7. Five-step checklist: how to reprice your stack
- Split the billing lines. Do not multiply the whole invoice by 1,100%. Cache-hit input, cache-miss input, and output have three different deltas. Peak hours are 9am–12pm and 2pm–6pm Beijing time.
- Move batch work off-peak. The official announcement asks for more flexible workload scheduling. Independent modeling of a heavy, mostly off-peak, half-cache-hit workload lands closer to a 1.8x bill, not 11x.
- Read the Qwen license, not the announcement thread. Personal projects and internal enterprise use are unaffected. The catch applies if you run a Model-as-a-Service or AI Work Assistant business that has earned over $50 million in any consecutive 12-month period. There is no geographic ban.
- Treat GLM-5.3 scores as vendor self-reports. The 4.6% → 28.3% Terminal-Bench jump is post-training only. No independent re-run is public. It still trails GPT-5.6 Sol (34.6%) and Claude Fable 5 (33.7%).
- Route by tier instead of "cheapest Chinese model." Off-peak DeepSeek V4-Pro still undercuts Claude Opus 5. Qwen3.8-Max international and GPT-5.6 Luna now undercut DeepSeek off-peak on at least one dimension. At peak, compare GMI Cloud and Novita against the official list.
8. Why this matters: two price wars running in parallel
Over roughly the past month, China's top labs have shipped major releases at a pace domestic financial media has started calling "three model updates a week" (一周三更) — DeepSeek, Alibaba, and Zhipu, plus Moonshot's Kimi K3 (open-weighted Jul 16, 2.8T parameters) and MiniMax H3 before them. Chinese coverage broadly frames this as Chinese open-weight releases "forcing a global repricing of the AI industry."
Meanwhile, US labs are running the opposite play at the consumer layer: OpenAI cut prices 80% on its cheapest tier (Jul 30) then made that model free and unlimited for all users a week later (Aug 6–7); Google shipped a coding-focused model at half the price of its three-week-old predecessor (Aug 13). Chinese labs open-weight flagships and introduce tiered, higher pricing on the compute-constrained top end; US labs race toward free and cheap at the consumer end. Both are real strategies; they optimize different parts of the funnel.
There is also a geopolitical layer worth naming carefully. Moonshot's Kimi K3 open-weighting in July already drew US security scrutiny; Alibaba choosing this specific window to open-weight a 2.4T flagship has been read by some analysts as a move to lock in international mindshare and a "technological parity" narrative before any potential regulatory tightening. That is an informed interpretation, not a confirmed fact — but it is part of the context that is hard to see if you only read English-language tech press, which has largely covered these releases as isolated product news rather than as a coordinated national pattern.
Meta's move is only a partial return to the Llama-era open-source posture. Muse Glimmer is a 30B distilled model, not Meta's real flagship. CEO Mark Zuckerberg has said open weights for the larger, closed Muse Spark 1.2 are coming "soon." If that happens, it would be the first US flagship-tier model released openly. As of this writing, that release has not happened; treat it as a stated intention, not a confirmed fact.
For operators, the practical response is not to pick a camp. It is to split the ledger: peak vs off-peak, cache hit vs miss, license thresholds, and vendor self-reports vs third-party reruns. Cloud APIs still absorb burst traffic and flagship A/B tests. Once official peak rates stop being the cheapest host and open checkpoints jump to trillion-parameter MoE, a laptop's unified memory and always-on agent stack usually fail first. A cleaner control path is to park long-context and 24/7 agent work on a metered remote Mac node instead of leaving every token on a single official peak-hour SKU.
9. FAQ
Is DeepSeek still cheaper than GPT-5.6 or Claude after the price hike?
Its off-peak rate is still cheaper than Claude Opus 5, but it is no longer the single cheapest option overall — OpenAI's GPT-5.6 Luna ($0.20/$1.20 per million tokens) and Alibaba's international Qwen3.8-Max pricing ($2/$6) now undercut DeepSeek's new off-peak rates on at least one dimension. DeepSeek is still relatively cheap for a frontier-class model, just not the outright cheapest anymore.
Can I use Alibaba's Qwen3.8-Max open weights for free in a commercial product?
Yes, for most use cases — personal projects and internal enterprise use are unaffected. The catch applies only if you are running a Model-as-a-Service or AI Work Assistant business that has earned over $50 million in any consecutive 12-month period; that tier requires a separate commercial license from Alibaba.
Is Qwen3.8-Max banned or restricted for US, EU, or UK users?
No. That claim circulated online but is false — the published license contains no geographic restriction of any kind. The restrictions are revenue-based, not tied to where you or your users are located.
What is actually different between GLM-5.3 and GLM-5.2?
Nothing at the base-model level — both use the same 743-billion-parameter foundation model. The performance gains (roughly 6x on Terminal-Bench 3.0) come entirely from scaling up reinforcement learning during post-training, with no retraining of the base model.
Will Meta actually open-source its flagship model, not just the smaller Muse Glimmer?
Not yet. Muse Glimmer is a 30B distilled model, not Meta's real flagship. Zuckerberg has said open weights for Muse Spark 1.2 are coming "soon." As of this writing, that release has not happened; treat it as a stated intention, not a confirmed fact.
10. Sources
- DeepSeek's official pricing announcement, cross-checked against Wall Street CN, IT Home, AIGC.cn, and V2EX community discussion
- Alibaba's official Qwen model repositories (Hugging Face / ModelScope) and South China Morning Post reporting on license terms
- Zhipu (Z.ai)'s official GLM-5.3 technical page, plus VentureBeat and StableLearn coverage
- Meta AI Research's official blog and VentureBeat coverage of Muse Glimmer
- Chinese financial outlets (Yicai / 第一财经, Sohu Finance) on the pacing and framing of China's open-weight release cycle
Pricing, license terms, and benchmark figures reflect publicly available information as of publication. Verify the latest official pricing and license terms before republishing, and note that details flagged above as unverified (domestic chip claims, the Cursor vulnerability report, and export-control rumors) have not been independently confirmed.
11. Close: after the peak-rate hike, you still need a place to run the control stack
Hosted APIs still work. The constraints are now explicit: peak list price is no longer guaranteed cheapest, the custom license recaptures large commercial users, and a 2.4T MoE checkpoint is not a laptop job. Windows and Linux GPU boxes are fine for throughput inference, but they are usually worse for Cursor toolchains, always-on OpenClaw, MLX quantization checks, and graphics-adjacent workflows. If you need one reproducible environment for DeepSeek off-peak routing, Qwen open-weight trials, and GLM-5.3 coding benches, park long-context and agent residency on a MACGPU remote Mac — SSH in, meter the node, keep the laptop for editing, and stop sending every token through a single official peak-hour SKU.