Chinese AI Models Claim 4 of Top 5 in Global Weekly Call Volume Rankings

210 Trillion Tokens: Competition Shifts from Benchmarks to Real Usage

Global AI large model weekly call volume reached 210 trillion tokens, up 8.25% week-over-week. Chinese models' weekly volume hit 116.18 trillion tokens (+16.71%), while US models fell to 13.98 trillion tokens (-32.4%). Chinese models have exceeded US volume for 28 consecutive weeks. Call volume acts as a "real usage heatmap" showing which models are integrated into products, code tools, and enterprise workflows — reflecting price, openness, channels, and ecosystem influence rather than raw quality or revenue.

Three takeaways for developers:

Model selection must weigh throughput, price, rate limits, and stability — not just single-response quality.

Open or open-access interfaces that lower integration barriers can rapidly expand call volume.

Real competition increasingly resembles infrastructure competition: whoever stably produces tokens at scale wins more workloads.

Call volume must be interpreted alongside task type. A model may gain volume from cheap batch processing, while another handles high-value code-agent tasks with lower volume but higher per-task value. Enterprises should track task success rate, per-call cost, human rework, and final business outcomes — not copy rankings for procurement.

Top Five: Four Chinese Models Occupy High Positions

Weekly global call volume ranking (top five, four Chinese models):

Rank 1: DeepSeek V4.1 Flash 1017 — 56 trillion tokens, +17% week-over-week

Rank 2: DeepSeek V4.1 Flash — 17.6 trillion tokens, +18% week-over-week

Rank 3: Zhipu GLM 5.3 Flash — 13.2 trillion tokens, -12% week-over-week

Rank 4: Tencent Hunyuan Hy4 — 12.8 trillion tokens, new entry

DeepSeek's two models combined exceed 73 trillion tokens, showing how low cost, high throughput, and developer ecosystem drive volume. GLM 5.3 Flash's decline despite holding third illustrates how version switches, channel changes, and task routing affect rankings. Notably, previous fifth-place Hy4 preview and ninth-place GPT-5.6 Luna dropped out — rankings shift rapidly with new versions, pricing, channels, and application integrations.

DeepSeek Leads for Third Week: Flash Route Prioritizes Scale

DeepSeek V4.1 Flash 1017 leads with 56 trillion tokens (+17%), while V4.1 Flash ranks second at 17.6 trillion (+18%). Flash models target high-frequency, low-latency, large-scale calls — classification, summarization, code explanation, log processing, batch extraction, and lightweight agent tasks — where unit cost and throughput matter more than peak single-response quality.

Enterprises can adopt tiered routing:

Simple high-frequency tasks -> Flash or low-cost models
Complex code and research -> High-reasoning models
High-risk production ops -> Model generates plan + human approval

This controls token spend and avoids routing all requests to the most expensive model. The real long-term metrics are task-level success rate and total cost, not weekly ranking position.

Hunyuan Hy4 Enters Top Four on Launch

Tencent Hunyuan Hy4 launched November 4 and immediately placed fourth with 12.8 trillion tokens. Hy4 totals ~770B parameters using a mixture-of-experts architecture and adds visual understanding for the first time. Compared to Hy4 preview, it strengthens instruction following, long-text comprehension, long-horizon agents, and real productivity tasks, plus Computer Use — combining command-line and GUI to execute complex tasks.

Computer Use's value goes beyond "mouse clicks": the model enters real software flows — opening pages, reading state, executing actions, verifying results, deciding next steps. Production use requires managing browser sessions, file permissions, network access, sensitive-operation confirmation, and full audit logs. Hy4's volume is still early; sustained entry into enterprise apps, code tools, and high-concurrency inference — plus stable pricing and service quality — will determine long-term ecosystem position.

Kimi K3.1 Pro Grows 91%: Long-Horizon Agents Consume Tokens

Moonshot's Kimi K3.1 Pro ranks ninth at 4.37 trillion tokens (+91% week-over-week). Launched October 24 with ~3 trillion parameters and a 1-million-token context window, it enhances code, 3D creation, financial research, and long-horizon agents. Long context and long-horizon agents significantly increase token consumption: models read more material, call more tools, retain more intermediate state, and retry on failure. Rapid volume growth may signal real workflow adoption or simply longer, costlier per-task execution.

Developers should break volume into a task ledger:

Task: Success rate, latency, human takeover rate

Model: Input, output, cache, and total tokens

Tools: Call count, failures, retries

Business: Rework, rollbacks, conversions, final delivery

Only by combining these can one judge whether volume growth means product penetration or runaway workflow cost.

US Model Volume Declines: What It Signals

US model weekly volume fell to 13.98 trillion tokens (-32.4%). A single week doesn't prove overall competitiveness decline — statistical scope, model routing, platform openness, version migrations, and regional differences all play roles. But from a product perspective, the drop highlights a trend: openness, low price, batch accessibility, and localized service are becoming key competitive variables. Even the most capable model struggles to enter daily workflows if integration cost is high, rate limits are strict, or APIs are unstable.

For model vendors, the next phase requires building complete token production systems: training infrastructure, inference clusters, channel distribution, developer tools, enterprise services, and cost observability — all essential. Domestic developers needing to integrate tools like Claude Code via compatible APIs can use Code80 (https://code.ai80.vip/home) for endpoint configuration, while managing model versions, key permissions, call budgets, data retention, and production audits on the project side.

FAQ

Does OpenRouter volume represent all global model calls?

No. It reflects only OpenRouter's statistical scope, suitable for trend observation, not total global volume across all platforms.

Does high token volume mean a stronger model?

Not necessarily. Volume is influenced by price, channels, concurrency, application integrations, and task types — not directly equivalent to model quality or revenue.

Why do Flash models easily achieve high volume?

They suit high-frequency, low-latency, batch, and lightweight agent tasks where large-scale applications prioritize unit cost, throughput, and stability.

Why did Hy4 enter top four immediately after launch?

Likely due to new capabilities, channel access, enterprise scenarios, and launch timing — but one week's data cannot determine long-term ranking; sustained volume and production adoption must be watched.

How should enterprises use the call volume ranking?

Treat it as a candidate signal, then evaluate models on real tasks for success rate, latency, token cost, rate limits, failure rate, and human rework — don't procure by ranking alone.

How can domestic developers access compatible APIs?

Use Code80 to learn integration methods, while independently managing model permissions, keys, data retention, call budgets, and audit logs.