What happened between Aug 12 and Aug 17, 2026
The three moves look contradictory until you treat them as one signal. Chinese labs are no longer competing only on sticker price. They are competing on who gets to set the price.
Jul 16 · Moonshot: Kimi K3 lands as a 2.8T open-weight model and draws US security scrutiny.
Aug 2–3 · Alibaba API: Qwen3.8-Max is previewed, then launched as a hosted endpoint.
Aug 10 · Meta contrast: Muse Glimmer ships at 30B under Apache 2.0. Meta teases open weights for the closed flagship Muse Spark 1.2.
Aug 12 · Weights go public: Alibaba publishes Qwen3.8-2.4T-A95B on Hugging Face and ModelScope. The same day, xAI ships Grok 4.6.
Aug 13–14 · Hike plus post-training: DeepSeek-V4-Pro goes GA and announces a price increase effective Aug 17. Google ships a cheaper Gemini 3.7 Flash. Zhipu follows with GLM-5.3 on GLM-5.2's 743B base.
Aug 17, 00:00 Beijing · New rates live: Zoom out and the US side is doing the opposite. On Jul 30 OpenAI cut GPT-5.6 Luna by 80%. On Aug 6–7 it made Luna the free default with unlimited text chats. One side raises prices and opens flagship weights. The other side goes free at the consumer layer.
Note: Chinese financial press frames this as open-weight releases forcing a global repricing. English-language tech coverage has mostly treated each drop as isolated product news. The pattern is easier to see if you keep both frames.
DeepSeek's new peak rates, Qwen3.8-Max specs, GLM-5.3 benches
The 1,100% headline is real and narrow. It applies to peak-hour cache-hit input on V4-Pro, the line that started closest to free. Output, which dominates most invoices, rose 350%. Peak hours are 9am–12pm and 2pm–6pm Beijing time. Figures below are per 1M tokens, from DeepSeek's announcement, cross-checked against Wall Street CN, IT Home, and V2EX.
| Billing item (per 1M tokens) | Old | 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% |
Independent cost modeling put a realistic heavy-usage month (about 84M tokens, mostly off-peak, half cache hits) closer to a 1.8x bill, not a 12x bill.
Qwen3.8-2.4T-A95B is the first time Alibaba has open-weighted a Max-tier model. Qwen3.5, 3.6, and 3.7 Max stayed API-only.
| Spec | Detail |
|---|---|
| Parameters | 2.4T total, 95B active per token (MoE, 512 experts, 10 routed + 1 shared) |
| Context | 262,144 tokens native on the open checkpoint, extendable to ~1.01M; hosted Max defaults to 1M |
| Cadence | Preview Aug 2 → API Aug 3 → open weights Aug 12 |
| International API | $2/M input, $6/M output |
| License | Not Apache 2.0. Custom Qwen3.8-Max License |
GLM-5.3 reuses GLM-5.2's 743B base. Zhipu reports the gains below. No independent third-party re-run has been published yet.
| 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 |
Caveat: GLM-5.3 still trails GPT-5.6 Sol (34.6%) and Claude Fable 5 (33.7%) on Terminal-Bench 3.0. Treat it as a top open-weight result, not an outright frontier win.
Is DeepSeek still the cheapest frontier model after the hike?
No. Off-peak V4-Pro is still well below Claude Opus 5. It is no longer the cheapest option on the table. Qwen3.8-Max's international API and OpenAI's Luna both undercut DeepSeek's new off-peak rate on at least one dimension.
Read the hike as a capacity problem, not a sudden pivot to premium branding. Flat, always-cheap pricing worked while GPUs kept up with demand. Once usage grew faster than supply, the price sheet had to show the constraint. "Encouraging more flexible workload scheduling" is corporate language for "peak-hour compute is scarce; shift the load yourself."
One detail most international coverage missed: at peak hours, DeepSeek's official API now prices above several resellers, including GMI Cloud and Novita. The assumption that the official endpoint is always the cheapest way to run DeepSeek is broken.
Alibaba's open-weighting is not a gift with no strings. The 2.4T checkpoint is free to download. The license is not Apache 2.0. Any Model-as-a-Service or AI Work Assistant business earning over $50 million in any 12-month window must negotiate a separate commercial license. Products with 100M+ monthly active users or $20M+ in monthly revenue must display the model name prominently. That is a different bet from Meta's Muse Glimmer, which ships under unrestricted Apache 2.0.
A rumor worth killing: claims that the license bans downloads from the US, EU, UK, or South Korea are false. The published text has no geographic clause.
GLM-5.3's story is the method, not the score. Same 743B base. No retraining. Terminal-Bench 3.0 moves from 4.6% to 28.3% by scaling reinforcement-learning environments in post-training. As pretraining returns flatten, post-training RL is becoming a cheaper lever than training a new foundation model. Mid-tier labs can close agentic and coding gaps without OpenAI-scale pretrain budgets.
| Model | Input / 1M tokens | Output / 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 via Alibaba's ratio) | ~$5.00 | ~$25.00 | No |
Beijing peak: 09:00-12:00 and 14:00-18:00 CST That is 01:00-04:00 and 06:00-10:00 UTC FX ref ≈ ¥7.15 / $1
"Chinese model = cheapest model" held for most of 2025 and early 2026. It is no longer a safe default.
Six steps to read the new DeepSeek bill and the Qwen license
Do not reprice your stack from a headline multiple. Split the line item, the clock, and the license trigger first.
Name the line item: 1,100%, 350%, and 200% are all correct. They describe cache-hit input, output, and cache-miss input. Output and cache-miss input move real invoices. Cache-hit input moves the headline.
Map your working day to Beijing peak: 9am–12pm and 2pm–6pm Beijing. US daytime is mostly off-peak on that clock. EU mornings overlap the second window. Batch jobs and eval loops should be scheduled explicitly.
Estimate from hit rate, not from the scariest row: A heavy off-peak mix with about half cache hits has been modeled near 1.8x. Peak plus low hit rate is where the 350% and 1,100% figures become real.
Check the Qwen revenue trips: Personal and internal use are fine. Separate commercial license if your MaaS or AI Work Assistant line exceeded $50M in any consecutive 12 months. Prominent model-name display if you have 100M+ MAU or $20M+ monthly revenue. No geo ban.
Treat GLM-5.3 as a post-training result: Same 743B base. Vendor-reported benches only. Keep the "open-weight first tier, not closed-frontier winner" label until a third party re-runs Terminal-Bench 3.0.
Split official API, resellers, and local weights: Official peak can now lose to GMI Cloud or Novita. Open weights move cost from tokens to host memory and uptime. If you need a machine that stays awake for local eval or agent loops, start with MESHLAUNCH Mac Mini rental pricing and the help center.
Unverified claims, two price wars, and three numbers you can quote
Keep these four items out of the "confirmed" column until primary sources catch up:
The 1,100% headline: Technically accurate, misleading without the tier. It is peak cache-hit input only.
Zhenwu M890 silicon: Several Chinese financial outlets say part of Qwen3.8-Max inference runs on Alibaba's in-house Zhenwu M890 chips and "Pangu AL128" supernodes. Alibaba has not published independent technical docs or third-party benches. Treat as vendor-adjacent and unverified.
A "serious" Cursor vulnerability: VentureBeat and Zhipu disclosed the claim. 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 remain speculative. No official announcement.
Domestic financial media has started calling the last month "three model updates a week." DeepSeek, Alibaba, and Zhipu sit on top of Moonshot's Kimi K3 (2.8T, Jul 16) and MiniMax H3. US labs ran the opposite consumer play: Luna down 80% on Jul 30, then free and unlimited on Aug 6–7; Gemini 3.7 Flash at half the price of its three-week-old predecessor on Aug 13.
There is also a geopolitical reading that English-language tech press largely skipped. Kimi K3 already drew US security scrutiny. Some analysts read Alibaba's 2.4T drop in this window as a bid to lock in international mindshare and a "technological parity" story before any tighter rules. That is an interpretation, not a confirmed fact. It is still part of the timing.
Three quote-ready figures: ① V4-Pro peak cache-hit input ¥0.025 → ¥0.30 (~1,100%); output ¥6.0 → ¥27.0 (350%). ② Qwen3.8-2.4T-A95B: 2.4T total / 95B active; international API $2 / $6. ③ GLM-5.3 Terminal-Bench 3.0: 4.6% → 28.3% on the same 743B base, vendor-reported.
Sources: DeepSeek's official pricing note, cross-checked with Wall Street CN, IT Home, AIGC.cn, and V2EX; Alibaba Qwen repos on Hugging Face / ModelScope and SCMP on license terms; Z.ai's GLM-5.3 page plus VentureBeat and StableLearn; Meta AI Research and VentureBeat on Muse Glimmer; Yicai and Sohu Finance on release pacing. Verify official pricing and license text before you republish.
Peak-hour APIs punish always-on eval loops. Open weights punish hosts that sleep or share ports. A laptop lid-close or a cheap VPS without Metal will distort both local benches and agent soak tests. For a production host that stays up for iOS CI/CD and AI agent automation, a MESHLAUNCH Mac Mini cloud rental is usually the cleaner option: dedicated Apple Silicon, 7×24, daily/weekly/monthly terms. See the pricing page and the help center.
Off-peak V4-Pro is still cheaper than Claude Opus 5. It is no longer the single cheapest option. GPT-5.6 Luna at $0.20 / $1.20 and Qwen3.8-Max international at $2 / $6 now undercut DeepSeek's new off-peak rates on at least one dimension. DeepSeek remains relatively cheap for a frontier-class model. It is just not the automatic cheapest pick.
Yes for most cases. Personal projects and internal enterprise use are unaffected. The extra license applies if you run a Model-as-a-Service or AI Work Assistant business that earned over $50 million in any consecutive 12-month period.
No. The published license has no geographic restriction. The limits are revenue-based, not location-based.
The base model is the same 743-billion-parameter foundation. The roughly 6x Terminal-Bench 3.0 jump comes from scaling reinforcement learning in post-training, with no retraining of the base.
Not yet. Muse Glimmer is a 30B distilled model. Zuckerberg has said Muse Spark 1.2 weights are coming "soon." Treat that as a stated intention. If you want a stable host for open-weight eval, start with the pricing page and the help center.