Tencent Hy4 Takes Second Place in a Usage Race DeepSeek Still Owns
The most useful thing about an LLM leaderboard is often the mistake it invites you to make. Tencent's Hy4 has climbed to second place on OpenRouter's weekly usage chart, processing 18.8 trillion tokens and passing the original DeepSeek V4.1 Flash at 16.9 trillion. That does not make Hy4 the second-best model in the world. It does make the launch too large to dismiss as another benchmark-shaped press release.
Hy4 was third with 16 trillion tokens in the previous snapshot. In data through November 10, it added 2.8 trillion and moved ahead of DeepSeek's older checkpoint, even though that model's usage grew 8% week over week. Tencent has turned a November 4 release into production-scale experimentation in under a week. Promotions helped—Hy4 is discounted by 50% on TokenHub through December—but discounts do not route prompts by themselves. Developers still have to decide that a model is worth integrating, testing, or putting behind an agent.
The larger story is that DeepSeek still owns this race. DeepSeek V4.1 Flash 1017 processed 53.8 trillion weekly tokens, 2.86 times Hy4 and more than the next three individual models combined. Add the original V4.1 Flash and DeepSeek's two variants account for 70.7 trillion tokens. This is not a changing of the guard. It is a widening field in which Chinese open-model vendors have captured the four highest-volume positions: two DeepSeek checkpoints, Tencent's Hy4, and Z.ai's GLM 5.3 Flash.
A usage chart measures gravity, not intelligence
OpenRouter is unusually explicit about the limitations of its ranking. It measures tokens, not accuracy, unique users, completed tasks, revenue, or customer satisfaction. Private requests are excluded. Variants are counted separately. Tokenizers differ, verbose models consume more units for the same job, and agent loops can turn one user request into millions of machine-generated tokens. A model can rise because it is popular, cheap, wordy, promoted, or especially suited to long-running automation. Usually the answer is some mixture of all five.
That distinction matters because DeepSeek says its 1017 checkpoint was deliberately tuned to produce shorter outputs and finish tasks faster. If the claim holds, a token leaderboard could understate its share of completed requests precisely because the model became more efficient. The unit being counted is an input to the product, not the product's outcome. Ranking models by tokens is a little like ranking databases by bytes read: useful evidence of real load, but a terrible substitute for knowing whether the query returned the right answer.
Still, 18.8 trillion tokens are observable behavior, and observable behavior deserves more weight than launch-day applause. Arena-style preference votes tell us which answer a user favored in a controlled comparison. Vendor benchmarks tell us how a checkpoint performs on a test chosen and implemented by its maker. Routing volume tells us which models developers are willing to run repeatedly, with money and latency attached. None is sufficient alone; together they describe capability, preference, and economic fitness.
The rest of the chart reinforces that split. GPT-6 Luna moved into fifth place at 7.04 trillion tokens, narrowly ahead of Step 5 Preview at 6.94 trillion. Claude Opus 5.5 is seventh at 6.76 trillion, while Claude Haiku 5.5 and Sonnet 5.5 sit nineteenth and twentieth. Yet OpenRouter's task-specific spend table places Opus first for classification, and premium closed models remain prominent on benchmark tables. The highest-volume model is not necessarily the model buyers reserve for the most valuable request.
Hy4 is competing on the shape of the workload
Tencent prices Hy4 at roughly $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens. Those numbers make long-context and repeated-prompt workflows plausible without requiring every request to justify premium-model economics. Hy4 also targets workloads that naturally generate substantial traffic: vision, GUI and command-line computer use, and training aimed at Blender and CAD. Tencent says the release improved 19.1% on GameDev and 9.8% on GameLogic compared with Hy4 Preview.
The company also reports a score of 3.04 across 579 financial tasks judged by 213 professionals, versus 2.97 for DeepSeek V4.1 Flash 1017 and 2.68 for both Kimi K3.1 Pro and GLM 5.3. It says internal search optimization reduced hallucinations from 12.95% to 7.2% and unsupported assumptions from 23.1% to 5.6%. These are specific and promising claims, but they remain vendor measurements until an independent team reproduces them. A decimal-rich benchmark is not automatically an objective one.
There is another caveat hiding inside the word “open.” Tencent said weights and system cards would arrive after launch, but its announcement did not yet link those artifacts when this data was collected. Hy4 may be accessible through an API, but accessibility is not auditability. Teams with licensing, data-governance, or self-hosting requirements should verify that the promised checkpoint, license, deployment recipe, and safety documentation actually exist before treating the model as an open-weight procurement option.
DeepSeek provides a useful contrast. V4.1 Flash 1017 is MIT-licensed, exposes a one-million-token context window, and recommends allowing at least 256,000 generated tokens for some tasks. Its model card reports Terminal-Bench 4.0 improving from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, and Automation-Bench from 54.8 to 60.5 over the prior checkpoint. More important for engineering budgets, DeepSeek says the model completes work faster with fewer output tokens. The enormous usage lead suggests that incremental checkpoint engineering can create more adoption than a new family name.
What engineering teams should test on Monday
Do not replace an incumbent because Hy4 has a silver medal. Put it into a controlled replay against the workload it claims to handle well. Start with 50 to 200 real, sanitized traces covering tool calls, image or chart interpretation, GUI navigation, long research sessions, and any CAD or Blender automation relevant to your product. Pin the exact model and provider; names such as “V4.1 Flash” already conceal checkpoints with materially different behavior.
Measure task success, not tokens in isolation. Record wall-clock time, input and output tokens, cache hits, tool-call failures, retries, human interventions, and cost per successful task. Evaluate concise and verbose models with the same completion rubric. A candidate that costs 20% less per token but needs two retries is not cheaper. A candidate that consumes fewer tokens but leaves an engineer repairing its output is not faster.
Run the evaluation in shadow mode before shifting live traffic, and keep a premium fallback for tasks with high error costs. For agent systems, cap loops and budgets independently of the model's advertised context window. For regulated or proprietary data, confirm where prompts are processed, whether providers retain them, and whether the exact route qualifies for your compliance requirements. Open weights do not make every hosted endpoint equivalent.
The actionable conclusion is not that US closed models have lost, nor that Hy4 has proved superior. The market is separating into two contests: maximum capability under a controlled test, and acceptable capability at a price that lets software call the model over and over. Hy4's rapid climb says the second contest now determines enormous volumes of real work. DeepSeek's lead says efficiency improvements compound into distribution. The next durable winner will not be the model with the prettiest launch chart; it will be the one teams can afford to keep inside the loop after the trial credits expire.
Sources: OpenRouter rankings, Tencent's Hy4 announcement, DeepSeek V4.1 Flash 1017 announcement, DeepSeek model card and benchmarks