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

The two-week window is easier to read as a table:

Zoom out further and the picture gets more interesting: 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. In other words, while Chinese labs were raising prices and opening up flagship weights, US labs were cutting prices and going free at the consumer layer, at the exact same time. That is not a coincidence; it is two sides of the same pricing fight. Background: Kimi K3 open weights and DeepSeek V4 GA peak/off-peak.

The pain points for buyers and builders:

• Headline multiples are not your bill. "11x", "1,100%", and "350%" are all true, but they map to different line items: cache-hit input, output, and cache-miss input.
• "Official is cheapest" is broken. At peak hours, DeepSeek's official API now prices above some resellers.
• Open weights is not Apache 2.0. Qwen3.8-Max is downloadable, but a custom license still fences high-revenue MaaS and AI work-assistant businesses.
• Geo-ban rumors outran the LICENSE file. Claims that Alibaba blocked US, EU, UK, and South Korea downloads are false. The published text has no territorial clause.

Three labs, three plays, one signal: China's AI labs are shifting from competing on price alone to competing on pricing power itself.

DeepSeek's new rates take effect Aug 17, 00:00 Beijing time. Peak hours are 9am–12pm and 2pm–6pm Beijing time.

The headline "1,100%" figure everyone quoted applies specifically to peak-hour cache-hit input pricing, the tier that started closest to free. Output pricing, which matters more for most real-world bills, rose 350%. 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 headline numbers. Cross-checked against DeepSeek's official announcement plus Wall Street CN, IT Home, and V2EX.

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.

1. DeepSeek: from flat-rate pricing to time-of-day pricing — this is a capacity problem, not a strategy pivot

The easiest misread of DeepSeek's move is "China's cheapest model finally caved to margin pressure." Look closer at the structure and it reads more like the opposite: a company making its compute constraints visible in the price sheet for the first time. DeepSeek's old 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" in the official announcement is corporate-speak for "our peak-hour compute is now scarce, please shift your load yourself."

One detail international coverage mostly missed: at peak hours, DeepSeek's own official API price is now higher than several third-party resellers (GMI Cloud, Novita, and others currently list V4 Pro below DeepSeek's new peak rate). The assumption that "the official API is always the cheapest way to run DeepSeek" — a core part of its reputation — has been broken for the first time.

2. Alibaba: open weights buy ecosystem goodwill; a custom license protects the revenue ceiling

Qwen3.8-Max's open-weighting is not a straightforward act of generosity. Alibaba did two things simultaneously: it published the full 2.4T-parameter checkpoint for free download, and it attached a custom license — not the permissive Apache 2.0 used for smaller Qwen models — 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, where "made-in-China model" still carries some hesitation among enterprise buyers), while keeping pricing leverage over the handful of companies actually 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. "Open weights" does not mean the same thing across these two releases.

One rumor worth killing explicitly: claims circulated online that Alibaba's license bans downloads from the US, EU, UK, and South Korea. That is false. The published license text contains no geographic or territorial clause of any kind. Primary sources: the Hugging Face repo and the LICENSE file.

3. GLM-5.3: no new base model, just a bigger post-training bet — and that is the real story

The most interesting fact about GLM-5.3 is not the score, it 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. This confirms a trend that has been building industry-wide for months: 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, and it is why mid-tier labs without OpenAI-scale compute budgets can still close the gap on agentic and coding benchmarks. Official page: Z.ai GLM-5.3.

TZ=Asia/Shanghai date +%H:%M
# peak if 09:00-12:00 or 14:00-18:00 Beijing
# shift batch / eval jobs off these windows
# do not quote 1100% as your real bill delta
curl -s https://api-docs.deepseek.com/quick_start/pricing

DeepSeek put a capacity shortage on the price sheet. Alibaba traded weights for ecosystem and kept leverage in the license. Zhipu showed post-training scale can stand alone as a lever.

RMB-to-USD conversion at about ¥7.15/$1, approximate. The short answer: no. Even after accounting for 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 on the table. 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.

DeepSeek is still relatively cheap for a frontier-class model. It is no longer the single cheapest option on the board.

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 pricing, the tier that started nearest to zero. Output pricing, the cost that dominates most real bills, rose 350%. Different outlets have quoted different tiers as if they were the whole story.
• Claims that Qwen3.8-Max runs on Alibaba's in-house Zhenwu M890 chips (reported by several Chinese financial outlets as evidence of a fully domestic-silicon inference stack) have not been independently confirmed by Alibaba's own technical documentation or third-party benchmarks. Treat this as vendor-adjacent, unverified reporting until confirmed.
• GLM-5.3's reported discovery of a "serious vulnerability" in Cursor comes from VentureBeat's reporting and Zhipu's own disclosure; specific technical details of the vulnerability have not been made public, so the claim should be read as a vendor-sourced, not independently audited, security finding.
• 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. Treat as background context, not established fact.

Why this matters: two price wars running in parallel

Place this in the bigger frame and a pattern emerges. 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" — a framing that is more assertive than most English-language coverage of the same events.

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). So while Chinese labs open-weight flagships and introduce tiered, higher pricing on the compute-constrained top end, US labs are racing toward free and cheap at the consumer end. Both are real strategies; they are just optimizing for 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 are only reading English-language tech press, which has largely covered these releases as isolated product news rather than as a coordinated national pattern.

Six-step checklist for builders and buyers:

1. Split the bill by line item. Recalculate cache-hit input, cache-miss input, and output separately. Do not treat "1,100%" as your real multiplier.
2. Map Beijing peak windows. Move batch Agent and eval jobs off 09:00–12:00 and 14:00–18:00 Beijing time.
3. Read the LICENSE file first. Use the Hugging Face / ModelScope text, not announcement threads claiming a US/EU/UK/KR geo-ban.
4. Check revenue and MAU triggers. MaaS or AI Work Assistant revenue over $50M in any 12 months needs a separate Alibaba license; 100M MAU or $20M monthly revenue needs prominent model-name display.
5. Compare before you lock a vendor. Put DeepSeek off-peak / peak next to Qwen international $2/$6, GPT-5.6 Luna $0.20/$1.20, and reseller quotes that now undercut official peak.
6. Pick a stable off-peak host. For 2.4T local pulls, long-context evals, or 7×24 Agent scheduling, prefer dedicated bare-metal Mac over oversold shared cloud.

# Aug 2026 China open-weight + pricing watchlist
1. split bill: cache-hit / cache-miss / output
2. shift jobs off Beijing 09-12 and 14-18
3. read Qwen3.8-Max LICENSE, not rumor threads
4. check $50M MaaS / 100M MAU / $20M monthly triggers
5. compare Luna $0.20/$1.20 vs Qwen $2/$6 vs DS off-peak
6. avoid shared-cloud for 2.4T / long-context evals
next: isolated Apple Silicon host

Citable figures (as of 2026-08-17):

• V4-Pro peak cache-hit input: ¥0.025 → ¥0.30, about 1,100% (the headline tier only)
• V4-Pro peak output: ¥6.0 → ¥27.0, 350% (the line that usually dominates real bills)
• Qwen3.8-2.4T-A95B: 2.4T total / 95B active, 262,144 native tokens, custom license
• GLM-5.3 Terminal-Bench 3.0: 4.6% → 28.3%, same 743B base, no pretraining rerun
• License triggers: $50M MaaS / AI Work Assistant revenue in any 12 months needs a separate license; 100M MAU or $20M monthly revenue needs prominent model-name display

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 (tied to how much money your service makes), 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. CEO Mark Zuckerberg has said open weights for the larger, closed Muse Spark 1.2 are coming "soon," which, if it happens, would make it 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.

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's 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.

Shared cloud hosts used for long-context evals or a 2.4T weight pull often show bandwidth jitter and oversubscription. Ad-hoc nodes used for off-peak Agent scheduling also drop long connections and rarely give you true isolation. Dumping high-memory inference onto an unstable shared instance shakes both the bill and the success rate. For a more stable production host for open-weight evals and off-peak Agent automation, JEXCLOUD multi-region bare-metal Mac is usually the better fit: dedicated Apple Silicon, root access, 7×24 uptime, monthly flexibility, about 120-second delivery. See the JEXCLOUD pricing page for nodes and rates.