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

Six pain points after the 1,100% headline

The headline number is loud.

The decision that actually stalls a team is usually one of these six gaps:

1. The hike percentages do not describe the same line item. “11x,” “over 1,100%,” and “350%” can all be true. They map to cache-hit input, output, and cache-miss input. Mix them and the budget is wrong.
2. Peak hours get ignored. The new peak window is 9am–12pm and 2pm–6pm Beijing time. The same workload can cost about 2x if it sits in that window.
3. “Official is cheapest” broke for the first time. At peak, DeepSeek’s official card now sits above some resellers, including current GMI Cloud and Novita quotes.
4. Open weights are not Apache 2.0. Qwen3.8-Max uses a custom license. MaaS or AI Work Assistant revenue over $50 million in any 12-month period needs a separate commercial license.
5. A geographic-ban rumor outran the license text. The official file has no US, EU, UK, or Korea download ban. Believing the rumor misreads compliance.
6. Vendor self-scores are not a cutover order. GLM-5.3’s Terminal-Bench jump has no third-party re-run. It still trails GPT-5.6 Sol and Claude Fable 5.

What happened: two weeks on one timeline

Zoom out and the contrast is the story.

On July 30, OpenAI cut GPT-5.6 Luna by 80%.

On August 6–7, Luna became the default for free users, with unlimited text chat.

Chinese labs spent the same window adding open weights and raising list prices.

US labs spent it cutting prices and giving tokens away.

That is one pricing contest with two fronts, not two unrelated news cycles.

Luna’s cut is unpacked in the GPT-5.6 price-cut brief.

The numbers: DeepSeek rate card, Qwen specs, GLM benchmarks

DeepSeek price card (effective Aug 17, 00:00 Beijing time)

Peak hours are 9am–12pm and 2pm–6pm Beijing time.

Unit: RMB per 1 million tokens.

Cache-hit input shows the largest percentage jump — about 12x / 1,100%+ on V4-Pro peak.

The absolute yuan amount on that line is still small.

Heavy callers feel output (+350%) and cache-miss input first.

Qwen3.8-2.4T-A95B (Qwen3.8-Max open weights)

This is Alibaba’s first Max-tier open-weight drop.

Qwen3.5, 3.6, and 3.7 Max stayed API-only.

GLM-5.3 vs GLM-5.2: same 743B base, post-training only

These scores are Zhipu self-reported.

No third-party re-run is public.

On Terminal-Bench 3.0, GLM-5.3 still trails GPT-5.6 Sol (34.6%) and Claude Fable 5 (33.7%).

That is a first-tier open-weight jump, not a clean sweep of the closed frontier.

Three strategies, three logics

1. DeepSeek: floor price to time-of-day pricing

This hike is easy to misread as “everyone else raised, so we raised.”

The structure looks more like a compute bill made visible.

DeepSeek had run a flat, low list price for a long time.

That model bought volume, then leaned on cache hits and off-peak scheduling to dilute GPU cost.

When call volume grew faster than GPU supply, the model stopped holding.

The notice’s line about “more flexible workload scheduling” is a plain translation: capacity is tight, so you shift the job.

One detail matters for procurement.

After the hike, official peak rates sit above some third-party resellers, including current GMI Cloud and Novita quotes.

“Official is the cheapest DeepSeek token” is no longer a safe default.

2. Alibaba: open weights for the ecosystem, a custom license for cash flow

Qwen3.8-Max is not a charity drop.

Alibaba did two things at once.

It published 2.4T-parameter weights.

It also attached a custom Qwen3.8-Max License instead of Apache 2.0.

MaaS or AI Work Assistant businesses that earn over $50 million in any 12-month period need a separate commercial license.

Products with 100 million+ MAU or $20 million+ monthly revenue must display the model name prominently.

The combo is coherent.

Open weights buy developer mindshare — especially overseas, where “Chinese models are not production-grade” still lingers.

Large cash-flow customers stay on a negotiated path.

That is a different “open” than Meta’s Muse Glimmer, which is Apache 2.0 with no extra commercial rider.

For a stricter open-weight baseline, see the Kimi K3 full open-weight release.

One rumor needs a hard stop.

Posts claimed the license bans downloads in the United States, the EU, the UK, and Korea.

That claim is false.

The official text has no geographic ban.

Rumors of that kind usually travel faster than a read of the file.

3. Zhipu GLM-5.3: keep the base, change the post-training recipe

The score is not the most interesting fact.

How the score was earned is.

The 743B base is identical to GLM-5.2.

There was no new pretrain.

Zhipu scaled reinforcement-learning environments in post-training.

Terminal-Bench 3.0 moved from 4.6% to 28.3% — about a 6x jump.

That matches a trend that has been getting clearer for about six months.

When pretrain scaling laws flatten, post-training RL scale becomes the next lever.

That lever is cheaper than training a new hundred-billion-plus base.

Smaller labs can close a gap with a better finish, not a bigger pretrain.

Head-to-head: is DeepSeek still the cheapest flagship?

FX is approximate at ¥7.15 / $1.

Use official cards for procurement.

Off-peak V4-Pro still undercuts Claude Opus 5 by a wide margin.

It is no longer the floor of the field.

Qwen3.8-Max international and OpenAI Luna both undercut DeepSeek’s off-peak list.

“Chinese = cheapest” is breaking into tiers.

Disputed and unverified details

• Hike-percentage mix-ups: “11x,” “over 1,100%,” and “350%” are all used in coverage. Each can be correct for a different billing line. Read the line item before you reuse the number.
• Zhenwu M890 / Pangu AL128 chip claims — unverified: Several Chinese financial outlets say part of Qwen3.8-Max inference runs on Alibaba’s Zhenwu M890 chips and “Pangu AL128” supernodes. The claim appears only in Chinese financial media. It is not independently confirmed by an Alibaba technical paper or a third-party benchmark.
• GLM-5.3 found a “serious Cursor vulnerability” — unverified: The claim comes from VentureBeat plus Zhipu. Details are not public. Treat it as vendor-sourced until an independent security lab publishes a write-up.
• China MOFCOM retaliatory export controls — unverified: Some reports say Beijing may prepare retaliatory AI / semiconductor export controls. There is no official confirmation. Treat it as speculative background, not a compliance fact.

Background: a two-front price war

These launches are not isolated product notes.

Over the past month, Chinese flagship labs have shipped on a near-weekly cadence.

Beside DeepSeek, Alibaba, and Zhipu sit Moonshot’s Kimi K3 (Jul 16, 2.8T open weights) and MiniMax H3.

Domestic financial press has framed the burst as Chinese open-weight models forcing a global reprice.

US labs took the other road.

OpenAI cut Luna 80% on July 30, then made the cheaper Luna the free-user default with unlimited text on August 6–7.

Google shipped Gemini 3.7 Flash at half price on August 13.

China is pairing open-weight flagships with a stepped price hike.

The US is pairing consumer-side free tiers with defensive cuts.

Both paths landed in the same August, and both rewrite how 2026 token prices get set.

Geopolitics is the other layer.

Kimi K3’s open weights already drew US security scrutiny.

Alibaba then open-weighted a 2.4T flagship in that same window.

Some analysts read that as locking in an international-equivalence story before any tighter controls.

That read is partly speculative.

It is still part of why this open-weight wave landed when it did.

Six-step runbook after the hike

Do not let the headline percentage set the budget.

Use these six steps to turn “will this cost more?” into a routing decision:

1. Mark the peak window first. Split current calls into Beijing 9am–12pm, 2pm–6pm, and the rest. Record each slice’s token share.
2. Split the bill by line item. Count cache-hit input, cache-miss input, and output separately. The 1,100% headline hits only the first line. Output at 350% is the main item on most bills.
3. Compare official, reseller, and substitutes. Peak official DeepSeek is not automatically cheapest. Check Qwen3.8-Max international ($2 / $6) and GPT-5.6 Luna ($0.20 / $1.20) on the same week.
4. Read the license file, not a screenshot. Individual and internal use are largely fine. A separate commercial license applies when MaaS or an AI Work Assistant earned over $50 million in the last 12 months.
5. Treat GLM-5.3 as first-tier open weight, not a full cutover. Self-scores have no third-party re-run. Terminal-Bench 3.0 still trails Sol and Fable 5. Shadow a slice of traffic. Do not flip 100% in one night.
6. Park off-peak batch and local weights on a node that stays up. A laptop lid closes the Beijing night window. 24/7 scheduling needs a dedicated macOS host for batch jobs, local weights, and audit logs.

# DeepSeek V4-Pro sketch: output line only, peak vs off-peak
# Rates from the 2026-08-17 card (CNY per 1M tokens)
PEAK_OUT, OFF_OUT = 27.0, 13.5

def monthly_output_bill(out_tokens_m, peak_share):
    peak = out_tokens_m * peak_share * PEAK_OUT
    off = out_tokens_m * (1 - peak_share) * OFF_OUT
    return round(peak + off, 2)

print(monthly_output_bill(10, 0.8))  # 80% peak → ¥243.0
print(monthly_output_bill(10, 0.2))  # 20% peak → ¥162.0

Three hard numbers worth citing

• ~1,100%: V4-Pro peak cache-hit input, from ¥0.025 to ¥0.30 per 1M tokens. Absolute yuan stay small. The headline is largest here.
• 2.4T / 95B: Qwen3.8-2.4T-A95B is 2.4T total and 95B active. Open weights are 262,144 tokens native, extendable to about 1.01M.
• 4.6% → 28.3%: On the same 743B base, GLM-5.3 is about a 6x Terminal-Bench 3.0 jump. It still trails Sol at 34.6% and Fable 5 at 33.7%.

Close: pricing power split. Always-on compute still has to sit somewhere.

Five August days stacked three facts on one rate card.

DeepSeek wrote a GPU shortage into peak and off-peak prices.

Alibaba traded open weights for an ecosystem and kept a custom license for bargaining power.

Zhipu showed that a post-training finish can land a model in the first open-weight tier.

Teams that run agents and local weights still hit the same three structural bottlenecks:

• Laptop sleep kills the off-peak window. Closing the lid throws away Beijing night rates. The headline hike then lands on the bill as-is.
• A shared dev machine is too wide. API keys, weight files, and batch queues on one sleeping laptop make peak/off-peak math and license audits both worse.
• There is no 24/7 scheduler in the lid. Off-peak batch, local open-weight inference, and call logs need a dedicated node, not a one-shot cloud function.

If you need off-peak scheduling, open-weight inference, and API-bill audit on a stable, isolatable production topology, MACCOME Mac cloud hosts give you real macOS, SSH handoff, and a controllable network boundary for 24/7 residency. Public plans are on Mac mini cloud rental rates.

Sources: DeepSeek official pricing, cross-checked with Wall Street CN, IT Home, AIGC.cn, and V2EX; Alibaba Hugging Face / ModelScope repos and SCMP on license terms; Z.ai GLM-5.3 page, VentureBeat, and StableLearn; Meta AI Research and VentureBeat on Muse Glimmer; Yicai and Sohu Finance on the Chinese release cadence. Figures reflect public information at publication. Recheck official notices for prices, license clauses, and benchmark scores before you commit a budget. Chip-hosting claims, the security-vulnerability report, and export-control rumors are labeled unverified in the body.