Tencent's Hy4 Flash Just Absorbed 46.6 Trillion Tokens — Don't Call It a Quality Ranking
The most important number on this week’s model leaderboard is not a rank. It is the 33.8 trillion tokens that appear to have moved into Tencent’s new Hy4 flash variant while the rolling window advanced. Hy4 flash now accounts for 46.6 trillion tokens on OpenRouter, nearly double the 22.8 trillion visible in the previous snapshot, while the older Hy4 sits at 13 trillion and is down 36% week over week. That is not proof that Tencent built the second-smartest model in the world. It is evidence that somebody, or several somebodies, found a reason to route an industrial quantity of work through the faster one.
That distinction matters because usage charts are easy to read as quality tables. OpenRouter’s ranking measures prompt and completion tokens processed through its public routing infrastructure. It does not count unique developers, requests, benchmark wins, customer retention, private traffic, or dollars spent. Models are listed as separate variants, and token totals are affected by output length and tokenization. A verbose model can literally score more usage for doing more talking.
Still, 46.6 trillion tokens is too large to dismiss as launch-week curiosity. Hy4 flash remains No. 2, behind DeepSeek V4.1 Flash 1017 at 62.1 trillion tokens and ahead of DeepSeek V4.1 Flash at 16 trillion. Tencent’s older Hy4 is fourth with 13 trillion. Z.ai’s GLM 5.3 Flash rounds out the top five at 10.9 trillion. The top nine models did not change position, but the apparently calm ranking conceals a substantial migration inside Tencent’s own product family.
A sibling swap is more useful than a podium finish
The best comparison for Hy4 flash is not DeepSeek’s leader. It is Hy4. When a fast variant rises sharply at the same time its sibling contracts, the plausible explanation is workload substitution: applications are moving traffic from the general model to a cheaper or lower-latency endpoint. The ranking alone cannot prove that causality—one very large application could dominate the total—but the paired movement is a stronger deployment signal than either model’s rank in isolation.
Hy4 flash processed 23.8 trillion tokens on its most recent complete day, up 14% day over day. That one day represents just over half of its trailing-week total. This makes the 46.6 trillion figure both impressive and fragile: if that daily pace persists, Hy4 flash could challenge DeepSeek for the top position; if it was a burst from a promotion, batch job, or single customer, the rolling chart will unwind just as quickly. The correct response is not to crown a winner. It is to add the model to an evaluation queue while the economic signal is fresh.
There is another reason to resist the podium narrative. OpenRouter’s task-spend view puts Claude Opus 5.5 first for implementation work with a 24.5% share, even though Opus is seventh in aggregate token volume at 7.63 trillion. Different workloads produce different winners. Bulk summarization, role-play, translation, coding agents, and structured extraction do not share the same tolerance for latency, verbosity, tool-call errors, or price. A global token chart compresses those trade-offs into one attractive but lossy number.
The lower half is moving faster than the market
The week’s dramatic rank changes mostly happened where the totals are tightly packed. MiniMax M3.1 Flash climbed three places to No. 10 with 3.67 trillion tokens. Kimi K3.1 Pro rose one position to No. 11 at 3.55 trillion, while GLM 5.3 slipped two places to No. 12 at 3.43 trillion. Baidu’s free ERNIE-5.5-Flash-Preview jumped three spots to No. 13 on 3.37 trillion tokens and a reported increase above 999%. NVIDIA’s free Nemotron 3 Ultra followed at 3.32 trillion.
Those ranks look decisive until you inspect the margins. Only 380 billion tokens separate No. 10 from No. 14, roughly one percent of Hy4 flash’s weekly volume. GPT-6.1 Sol fell four places to No. 15 with 3.29 trillion tokens, but it trails No. 10 by just 380 billion. DeepSeek V4 Flash 0731 dropped five places to No. 19 at 2.88 trillion. In this band, a modest routing change can produce a dramatic arrow without establishing a durable trend.
Free access also complicates interpretation. ERNIE’s breakout and Nemotron’s presence in the top 20 show that price is part of the benchmark, whether the chart labels it that way or not. Free previews are useful distribution tools: they invite experimentation, seed integrations, and generate public usage. But they test demand under subsidy, not necessarily willingness to pay. The question for Baidu is not whether a free model can attract tokens. It is whether those workloads remain when the preview ends, rate limits tighten, or a paid meter appears.
No model entered or left the top 20 in this snapshot. That makes the internal reshuffling more informative: developers are reallocating traffic among an established candidate set rather than discovering an entirely new one. It also argues for patience. A three-day persistence test is more useful than reacting to a single daily arrow, especially below the top nine.
Turn the leaderboard into a test plan
Platform teams should use this chart for candidate generation, not procurement. Hy4 flash has earned a controlled trial on high-volume, latency-sensitive paths. Start with a representative sample of real production traces and replay them against the current model and the challenger. Grade task completion, not prose preference. For tool-using systems, measure valid tool selection, argument correctness, retry frequency, and the rate of malformed structured output. For retrieval workloads, score citation support and long-context recall. For customer-facing generation, track refusal behavior, factual errors, and tail latency.
Run the comparison in shadow mode before shifting live traffic, then segment the results by task. A single aggregate win rate can hide a model that is excellent at extraction and unreliable at multi-step tool use. Record input tokens, output tokens, wall-clock latency, and actual cost per successful task. “Per successful task” is the denominator that matters: a cheap call that needs two retries is not cheap, and a terse model may beat a verbose one even when their advertised token prices look similar.
Teams should also put routing changes behind an observable switch. Canary a small share, establish rollback thresholds, and watch output length as closely as error rate. Because OpenRouter counts both prompt and completion tokens, a ranking surge may reward behavior that would inflate an application’s bill. If Hy4 flash wins on your traces, expand it by workload rather than declaring it the new global default.
One source is conspicuously absent today: Arena’s text leaderboard could not be verified. The dedicated collection returned no data and the fallback hit Cloudflare, so there is no defensible head-to-head preference ranking to pair with the usage numbers. That missing evidence should narrow the claim, not invite a guess. Public preference tests and routed usage answer different questions anyway; neither substitutes for an evaluation built from a team’s own failure modes.
The editorial verdict is straightforward. Hy4 flash is not “the No. 2 LLM.” It is the second-largest public token sink on OpenRouter this week, and the simultaneous decline of Hy4 makes that a credible migration signal. Engineers should investigate it now, test it on their own traces, and ignore the temptation to convert traffic into intelligence. The next meaningful milestone is not whether Tencent reaches No. 1. It is whether the traffic survives the rolling window—and whether independent teams can reproduce the speed, cost, and reliability trade-off that apparently attracted it.
Sources: OpenRouter AI Model Rankings, OpenRouter models catalog, OpenRouter documentation