Why Did DeepSeek Just Raise API Prices by Up to 1,100%
— Right After China's Open-Weight AI Blitz?

In a five-day window, teams that buy Chinese frontier APIs or download open weights hit three moves that look contradictory on the surface. DeepSeek raised API prices by as much as 1,100% on certain tiers. Alibaba, the same week, open-weighted a 2.4-trillion-parameter flagship it had never released before. Zhipu AI shipped GLM-5.3, boosting coding benchmarks by roughly 6x on the exact same base model — no retraining. This piece is for engineers who need to recost a bill, read a license, and pick a stack. The shared signal: China's labs are shifting from competing on price alone to competing on pricing power.

SECTION 01 Five ways teams misread the DeepSeek hike

The story is not a shortage of headlines. It is a shortage of clean billing dimensions.

• Treating "1,100%" as the whole invoice: that figure is peak-hour cache-hit input on V4-Pro. Output rose 350%. Cache-miss input rose 200%.
• Assuming the official API is always cheapest: at peak hours, DeepSeek's own list price now sits above several resellers (GMI Cloud, Novita, and others still quote V4 Pro below the new official peak).
• Reading Alibaba's drop as Apache 2.0 charity: the weights are downloadable, but a Model-as-a-Service or AI Work Assistant business over $50 million in any 12-month period must negotiate a separate commercial license.
• Repeating the geo-ban rumor: claims that the US, EU, UK, and South Korea cannot download the checkpoint are false. The published license has no territorial clause.
• Calling GLM-5.3 a new foundation model: it reuses GLM-5.2's 743B base. Gains come from scaled post-training RL. The scores are vendor-reported; no independent re-run is public yet.

Three labs, three plays, one signal: China's AI labs are competing on pricing power, not just price.

SECTION 02 Timeline: what shipped, and when US labs cut prices

Zoom out. On Jul 30, OpenAI cut its cheapest tier, GPT-5.6 Luna, by 80%. On Aug 6–7 it made Luna the free default with unlimited text chats. While Chinese labs raised prices and opened flagship weights, US labs cut prices and went free at the consumer layer. That is two sides of the same fight. Earlier product notes: Qwen3.8-Max launch and DeepSeek V4 GA pricing.

SECTION 03 The numbers: DeepSeek rates, Qwen specs, GLM-5.3 benches

New DeepSeek rates take effect Aug 17, 00:00 Beijing time. Peak hours are 9am–12pm and 2pm–6pm Beijing time. The 1,100% headline applies to peak cache-hit input — the tier that started closest to free. Output, which dominates most real bills, rose 350%. Independent cost modeling found a realistic heavy-usage workload (roughly 84M tokens/month, mostly off-peak, half cache hits) sees a bill increase closer to 1.8x.

These are Zhipu's own numbers. 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.

SECTION 04 Three strategies: time-of-day pricing, licensed weights, post-training

DeepSeek: from flat-rate cheap to time-of-day pricing — a capacity problem, not a brand pivot. The easy misread is "China's cheapest model finally caved." The structure reads more like compute constraints made visible. Flat, always-cheap pricing worked as acquisition while GPU supply kept up. Once usage grew exponentially and capacity did not, something had to become explicit. "Encouraging more flexible workload scheduling" is corporate-speak for scarce peak-hour compute. At peak, the official API is no longer the cheapest way to run DeepSeek. That assumption just broke.

Alibaba: open weights buy mindshare; a custom license protects the revenue ceiling. Publishing the 2.4T checkpoint and attaching a custom license are one move. 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. Products with 100M+ monthly active users or $20M+ in monthly revenue must display the model name. The bet: win international developers, keep leverage over anyone who can sell inference on top. That is a different bet than Meta's Muse Glimmer under unrestricted Apache 2.0. The geo-ban rumor is false — check the LICENSE file, not the announcement thread.

GLM-5.3: no new base, a bigger post-training bet. Same 743B base as GLM-5.2, no retraining, and a roughly 6x jump on Terminal-Bench 3.0 (4.6% → 28.3%) from scaling RL environments. As pretraining scaling laws flatten, post-training RL is becoming an independent lever with a lower cost floor than a new foundation model. Mid-tier labs can still close the gap on agentic and coding benches without OpenAI-scale pretrain budgets.

SECTION 05 Is DeepSeek still the cheapest frontier model? A 6-step audit

RMB-to-USD at about ¥7.15/$1, approximate. Off-peak V4-Pro is still well below Claude Opus 5, but it is no longer the outright cheapest option. Qwen3.8-Max international pricing and OpenAI's Luna both undercut DeepSeek's off-peak rate. "Chinese model = cheapest model" held through most of 2025 and early 2026. It is not a safe assumption now. Context on the US cut: GPT-5.6 Luna down 80%.

If you have to ship against this week, audit the stack in six steps instead of multiplying the headline:

1. Split the invoice: break the last 30 days into cache-hit input, cache-miss input, and output. Apply the new card to each line. Do not multiply the whole bill by 1,100%.
2. Mark Beijing peak hours: 9am–12pm and 2pm–6pm Beijing time. The announcement asked for more flexible scheduling for a reason.
3. Compare official vs reseller: at peak, check whether GMI Cloud, Novita, or another reseller still undercuts DeepSeek's official peak rate.
4. Read the LICENSE file: before you download Qwen3.8-2.4T-A95B, confirm the $50M / 100M MAU / $20M monthly-revenue thresholds. Ignore geo-ban threads.
5. Treat GLM-5.3 as a post-training delta: same base as 5.2, vendor benches only. Re-run your own terminal and SWE tasks before you swap.
6. Route by layer: cheap consumer text to Luna; open weights and ecosystem to Qwen; peak-shiftable frontier inference to DeepSeek off-peak; long-horizon agents and iOS CI to a real Mac, not a hypervisor.

peak = {"in_miss": 9.0, "in_hit": 0.30, "out": 27.0}
off  = {"in_miss": 4.5, "in_hit": 0.15, "out": 13.5}

tokens_m = {"in_miss": 40, "in_hit": 40, "out": 4}
def bill(rate, t):
    return t["in_miss"]*rate["in_miss"] + t["in_hit"]*rate["in_hit"] + t["out"]*rate["out"]

print("off-peak CNY", round(bill(off, tokens_m), 2))
print("peak CNY", round(bill(peak, tokens_m), 2))

SECTION 06 What is unverified, why two price wars, and which sources to reopen

• The 1,100% headline is accurate and incomplete: it is peak cache-hit input only. Output — the line that dominates most bills — rose 350%.
• Zhenwu M890 / Pangu AL128 claims: several Chinese financial outlets say part of Qwen3.8-Max inference runs on Alibaba's in-house chips. Alibaba has not published independent technical docs or third-party benches. Treat as unverified.
• GLM-5.3 and a "serious Cursor vulnerability": VentureBeat plus Zhipu's own disclosure. Technical details are not public. Read it as vendor-sourced, not independently audited.
• Retaliatory export controls: reports that China's Ministry of Commerce may prepare AI/semiconductor countermeasures are speculative media, not an official announcement.

Over the past month, China's top labs have shipped at a pace domestic financial media calls "three model updates a week" (一周三更) — DeepSeek, Alibaba, Zhipu, plus Moonshot's Kimi K3 (open-weighted Jul 16, 2.8T) and MiniMax H3. Chinese coverage frames this as open-weight releases forcing a global repricing. US labs ran the opposite play at the consumer layer: OpenAI's 80% Luna cut, then free unlimited text; Google's half-price Gemini 3.7 Flash on Aug 13. Flagship open weights plus tiered, higher prices at the compute-constrained top; free and cheap at the consumer end. Both are real. They optimize different parts of the funnel.

There is also a geopolitical layer. Kimi K3 already drew US security scrutiny. Some analysts read Alibaba's 2.4T drop in this window as locking in international mindshare and a "technological parity" narrative before any regulatory tightening. That is an informed interpretation, not a confirmed fact — and easy to miss if you only read English product posts.

Official and cross-checked sources. Re-open these pages before you republish prices or license terms:

DeepSeek official API pricing (peak / off-peak card)

Alibaba Cloud Community: Qwen3.8-Max launch note

Hugging Face: Qwen3.8-2.4T-A95B checkpoint

Reuters / MarketScreener: DeepSeek raises V4 API prices

MarkTechPost: GLM-5.3 same-base post-training

Self-hosting a 2.4T MoE or wiring an agent into Xcode and iOS CI stacks hypervisor tax on top of a token bill that just went up. For zero-loss native compute, stable iOS CI/CD, and 24/7 agent automation, a MACNOX cloud physical node is usually the better production path: genuine Apple hardware, full Root, no hypervisor, day / week / month billing. Share context: OpenRouter July rankings.

SECTION 07 FAQ