Tencent Hy4 Tops OpenRouter’s Weekly Ranking: Why Are Developers Using More Agent Models?

TL;DR

OpenRouter’s official ranking shows Tencent Hy4 preview processed 14.7 trillion tokens in the week ending September 6, 2026, ranking first and rising 379% from the previous period. This is a strong adoption signal on the OpenRouter distribution platform, but it does not prove that Hy4 is the world’s most-used model or that token volume alone proves model quality.

What happened

OpenRouter ranks models by the number of tokens processed through its API. The visible leaders through September 6 also include GPT-5.6 Luna, GLM 5.3 Flash, DeepSeek V4 Flash, MiniMax M3, and Xiaomi MiMo-V2.5.

OpenRouter’s model page lists Hy4 preview as a mixture-of-experts model with 770B total parameters, 49B active parameters, roughly 1.05M context, and an August 28, 2026 release date. Its stated positioning focuses on coding agents, complex tool use, and sustained multi-step productivity work.

What the data does and does not show

1. Platform usage is not global market share

The OpenRouter ranking covers aggregated public traffic routed through OpenRouter’s API. Direct provider APIs, private enterprise deployments, subscription products, and undisclosed applications are not fully represented. “Hy4 topped OpenRouter” is a platform fact; it should not be rewritten as “Hy4 is now the world’s number-one model.”

2. Token volume is not model quality

Usage can be affected by price, context length, provider availability, routing, application traffic, and workload mix. The ranking is useful for asking what developers are calling through this platform, but it cannot answer which model is smartest by itself.

3. Agent workloads are an important lens

Hy4 is positioned for coding agents, tool use, and long-context workflows. The more useful follow-up questions are where it is used, what it costs, where it fails, and whether it can deliver stable results in real workflows.

Implications for AI practitioners and business opportunities

• Model selection services: record task type, version, routing, input and output tokens, price, latency, failures, and privacy conditions when comparing models.
• Agent cost governance: inspect actual workflow logs instead of inferring savings from a leaderboard; cache hit rates and savings must be measured from real logs.
• Content opportunity: a follow-up tutorial can compare Hy4, DeepSeek, GLM, and Claude on reproducible coding-agent tasks, but only after a real test record exists. This article does not claim to have run that test.

What changed from the previous article

The earlier article focused on the May 2026 Hy3 ranking story and included historical cost and cache inferences. This revision uses the official September 6 OpenRouter ranking and current Hy4 model page, removing older figures and performance conclusions that were not re-verified in this update.

Conclusion and limits

The defensible conclusion is that Hy4 preview has become the top model by weekly token usage on OpenRouter, a meaningful developer-adoption signal. It is not independent proof of global market share or model quality. A tutorial should come next only after a reproducible test records versions, prompts, routing, costs, outputs, and failures.

Sources

• OpenRouter official model rankings
• OpenRouter: Tencent Hy4 preview
• NBD: September 7, 2026 model usage report