The industrial history of gateway wars is a recurring cycle of "overcapacity, channels reign supreme."

In the PC era, operating systems and browsers won; in the mobile era, App Store icon placement prevailed. Each round follows the same script: an arms race first levels capabilities, then competition shifts upward to "the path users take to access those capabilities." Whoever offers the shortest path and highest frequency of use captures bargaining power across the entire industrial chain (industrial chain).

In AI’s first round, the model itself was the gateway. ChatGPT’s minimalist dialog box was a dimensional reduction strike against search bars—a result of generational capability gaps that allowed extreme simplicity. The 2026 landscape differs: flagship models now lack generational gaps, open-source and domestic models have commoditized "sufficient capabilities," and a 5-point Review List (benchmark) gap fails to secure a second user click.

The competition thus returns to an updated version of an old question: When capabilities are equalized, which shell gets opened?

The previous AI competition pitted models against each other; this round pits shells against each other. This week, three companies from China and the US provided three answers: instant software generation, API call habits, and high-frequency native gateways—the three battlefields of this war.

01 The Calculus of GPT-6’s Free Tier: Software’s New Distribution Channel

Why would OpenAI give its most powerful model away for free to 1.2 billion people? The answer lies in Intelligent UI’s technical essence.

OpenAI describes it as "a native, streamable component library paired with a compiler that works synchronously during generation." In plain terms, ChatGPT gains the ability to "compile" mini-software on the fly during conversations: need to split bills? Generate a splitter instantly. Need to disassemble a bicycle? Generate a clickable diagram immediately. No downloads, installations, or redirects—use and discard.

The industrial implications run deeper than experience upgrades.

Mobile internet’s organizational logic trapped functions within fixed App interfaces, with developers competing for desktop icon placement. Intelligent UI undermines this foundation: when users grow accustomed to "having AI generate whatever they need," ChatGPT becomes software’s new distribution channel—and distribution channels never monetize through subscriptions. Browsers profit from default positions, app stores from commissions, operating systems from ecosystem taxes. Free isn’t generosity; it’s trading subscription market cap for gateway market cap: Gateway Value > Subscription Value.

Supporting moves all serve this logic. "Think While Speaking" reduces first-response wait times by 44% because wait time is the first law of distribution channels—no one forms habits around loading screens. 1.2 billion weekly active users form the distribution base: the same capability change, embedded in 1.2 billion people’s default dialog boxes versus a downloadable new App, represents two orders of magnitude in warfare.

The pressure list becomes clear: single-function tool Apps and shell-type AI Apps. The former get replaced by "on-the-fly generation," the latter by "official natives." The rationale for installing an App for a single function is vanishing.

02 The Calculus of Haiku’s Price Cut: API Gateway Melee

Anthropic lacks 1.2 billion weekly active users, so it targets a different gateway: enterprise developers’ default call habits. The API console is Anthropic’s version of a "desktop icon."

For inputs under 100,000 Tokens: $0.10 per million inputs, $0.50 per million outputs—down 90% from the previous generation’s $1 and $5, with ~90% of requests falling under this threshold. A new tokenizer increases Token consumption by ~25%, but average costs still drop ~75%. These numbers precisely target enterprise deployments’ largest, most repetitive call layers.

But this move reads more like reactive warfare.

GPT-6 Luna’s base input price already sits at $0.10—Haiku 5.5 "matched" rather than "led" on pricing. Meanwhile, domestic small models undercut further: Qwen3.7 Flash charges $0.03 for short- file (tier) inputs, while GLM-5.3-Flash outscores Haiku by one point. The small model market shifted overnight from "cost-performance wars" to "penny pricing."

Anthropic’s confidence comes from structural advantages: in the agent era, call volumes amplify through "sub-agents." Haiku 5.5’s official role is to support Opus 5.5 and Sonnet 5.5—every large model orchestration task spawns multiple small model calls. Large models sell intelligence; small models sell throughput. The pricing economics thus hold: trade price for call volume, trade call volume for revenue scale.

This narrative serves another urgent purpose: IPO preparations. The Claude 5.5 family released three models (Opus, Sonnet, Haiku) in a month, with Reuters confirming IPO plans. The prospectus needs a curve showing revenue sharply rising with call volume.

03 China’s Parallel War: WeChat and Mobile Assistants

China’s trigger pulled in early September.

On September 8, Meta launched Muse, an "action-oriented" agent that handles emails, bookings, and cross-platform tasks, gaining overseas traction. Within three weeks, all Chinese giants entered the fray: Manus’s Cue personal assistant (September 28), Alibaba’s QianWen Personal Agent for 300 million users at Cloud Town (September 22). ByteDance and Tencent accelerated post-Muse.

ByteDance’s "Spell" project had early foreshadowing: project initiation (initiated) in April 2025, building on Doubao Mobile Assistant’s personal agent direction, led by the mobile assistant team. Reportedly named "Xiaodou," it will launch as a standalone App; Doubao Mobile Assistant’s consumer version just landed on Nubia models in September. ByteDance’s calculus is clear: in China’s fragmented Android ecosystem, system-level assistants occupy a more fundamental shell than any App—every screen wake, every long-press, is a gateway.

Tencent’s playbook uses a different shell. LightVela positions as a "cloud-based personal assistant" integrated with WeChat, processing daily requests 24/7. WeChat is China’s highest-frequency digital gateway by far; an agent integrated into WeChat opens a service counter inside 1.4 billion people’s contact lists. Tencent also primes its models: Hunyuan Hy3 boosted task resolution rates from 72% to 90% after integrating with WorkBuddy; on October 12, Step 5 Preview from Step.cn also integrated with Tencent WorkBuddy—a third-party model placing its flagship into another’s gateway speaks volumes about the landscape.

The US and Chinese battlefields are structurally identical: model capabilities must parasitize high-frequency gateways to deliver value. The shell determines delivery efficiency.

The difference lies in pieces: the US lacks a gateway harder than ChatGPT, so OpenAI upgrades its dialog box into a distribution platform. China lacks no gateways but needs to imbue WeChat and mobile assistants with "acting on your behalf" capabilities, so ByteDance and Tencent flood models into gateways.

04 Conclusion: No Seat for Bystanders

The real spectacle now shifts to retention rates after free trials end.

Four observable coordinates emerge:

1. Revisit rates for Intelligent UI tools in ChatGPT—proof of whether distribution channels materialize;

2. Haiku 5.5’s call volume elasticity and revenue curve post-price cut—the IPO prospectus will tell the truth;

3. Opening data for Doubao’s "Xiaodou" standalone App and WeChat’s LightVela—testing China’s two shells;

4. Paid conversion rates after Step 5’s two-week free trial—representing third-party models’ ROI for surrendering to gateways.

Historical patterns of moat relocation dictate: after iteration, competition at the previous layer quickly becomes uncontroversial. Models will resemble utilities—stable, qualified, always available. The shell, however, will bear a family’s name, a family’s address. The gateway war rages precisely over whose address gets remembered.

Henceforth, the most frequently opened dialog box in global users’ phones will quietly rewrite software distribution’s underlying protocols. This war admits no bystanders—everyone’s opening habits are the battlefield itself.