Yesterday we explained why DeepSeek-V4-Pro changed the standard for what Baidu AI answers cite. Today comes the follow-up signal: DeepSeek is now charging for it — seriously. On August 17, a peak/off-peak pricing structure takes effect that raises peak V4-Pro output prices from ¥6 to ¥27 per million tokens, a 350% increase. Cache-hit prices jump 12x. Even off-peak hours cost 2.25x more.
For overseas brands, this is not a supply-side story about one Chinese AI company. It is the clearest evidence yet that China's AI reference layer is becoming a real, monetized market — which changes the economics of who gets cited, and why.
💸 The Price Story — What Changes on August 17
The new pricing, effective 00:00 August 17, is a three-tier structure:
| Tier | V4-Pro Output (per M tokens) | Change |
|---|---|---|
| Peak hours | ¥27 | +350% (from ¥6) |
| Off-peak hours | ¥13.5 | +225% (2.25x) |
| Cache-hit output | ¥0.30 | 12x (from ¥0.025) |
This is not a small adjustment. A 350% jump on the flagship model's output price, paired with a 12x cache-hit increase, signals a deliberate strategy shift: stop subsidizing scale, start monetizing value. The "price butcher" that undercut every global competitor by 57x is putting the knife down — at least for the model that now powers the frontier tier of Chinese AI search.
🔪 Why the Price Butcher Turned
Three forces explain the reversal, and none of them are random:
- Supply-demand imbalance. On August 1, V4-Flash processed 8 trillion tokens in a single day. By August 4, capacity was so stretched the service was nearly unavailable. At ¥6 per million output tokens, demand grows faster than compute can be added.
- The whole industry is repricing. In Q2 2026, average Chinese large-model API prices rose 48% for input and 80% for output. DeepSeek is not moving alone — the entire domestic API market has shifted from customer acquisition to monetization.
- The gap with the US is closing. China-US API price gaps have narrowed from extreme to 19% (input) / 14% (output). The arbitrage window that made China's models a pure cost play is closing, which forces buyers to evaluate on capability, not just price.
🏗️ What Commercialization Means for the AI Layer
Some overseas marketers might read this as "AI search in China is getting more expensive, so the opportunity is shrinking." The opposite is closer to the truth:
- A monetized citation layer is a stable citation layer. Free or subsidized models create chaos — suppliers rotate, capacity fails, quality varies. Paid infrastructure behaves like infrastructure: reliable, documented, investable.
- Enterprise-grade selection rewards existing assets. When AI platforms cost more to run, their answers favor sources that are cheap to verify: structured, authoritative, cross-checked content. That is exactly the asset stack (Baidu Baike, ICP-registered sites, news sources, Baijiahao) we keep describing.
- Capability is now independently verified. The same week saw DeepSeek Harness open-sourced under MIT and V4-Pro rated in the global top tier by securities research (SWE-bench 96.4%, second worldwide). Commercial pricing plus third-party validation means the engine is being treated as a serious product — and its citation behavior will be treated seriously too.
🌏 What This Means for Overseas Brands
For an overseas brand entering China, the pricing shift changes three calculations:
- Your citation assets appreciate, not depreciate. If the AI layer is becoming a paid, stable market, then the Baike pages, verified news mentions, and ICP-registered site you build are durable infrastructure — not a temporary SEO trick.
- Competition gets more rational. When price alone stops being a differentiator, domestic brands can no longer buy their way into AI answers with mass-produced soft articles. The quality bar that overseas brands can meet through systematic asset building rises relative to volume-based tactics.
- Baidu's cost structure shifts with the market. Baidu's Wenxin assistant already runs DeepSeek-V4 (integrated August 7). As model costs rise, the economics of AI answers favor efficient, verifiable sources even more — raising the value of being citable rather than merely present.
📌 Three Steps Before the Monetized Market Matures
- Build the verifiable Chinese-language asset stack — Baike, ICP site, news citations, active Baijiahao.
- Structure data for the agent layer — monetized models will "call" machine-readable, verified sources.
- Act before the window closes — first movers on a category's default answer keep it.
🏢 The BPP Perspective
The narrative arc of the last two days completes itself: yesterday, the engine got stronger (1M context, agent scores at frontier level). Today, the engine started charging real money. The message for overseas brands is that China's AI reference layer is now a real market with real economics — and real markets reward assets, not shortcuts.
For overseas brands, the sequence remains the same, with one new justification:
- Build the verifiable Chinese-language asset stack — Baike entry, ICP-registered site, news citations, active Baijiahao. These are the sources a paid, cost-conscious AI engine will prefer to cite.
- Structure your data for the agent layer — as models become monetized products, their "call" behavior (not just citation) will follow verified, machine-readable sources.
- Act before the window closes — the same reasons that make AI answers more reliable also make them more contested. Domestic brands are building assets at scale; the first mover on a category's default answer keeps it.
Every step requires a Chinese entity — the account, the Baike entry, the ICP filing, the payment rails. That is the structural wall BPP exists to bridge. If you want to be citable in China's AI answers before the monetized market matures further, the time to build is now — and the bridge is already in place.