The LLM Leaderboard Froze at the Top. The Useful Signal Is Underneath
The most useful thing an LLM leaderboard can do is refuse to move. OpenRouter’s five busiest models held their positions this week, which strips away the launch-day drama and leaves a harder, more practical question: what does sustained usage actually tell an engineering team? Not which model is “best.” It tells us which endpoints developers are willing to keep sending work after the first round of curiosity wears off.
That distinction matters because the numbers are enormous enough to invite lazy conclusions. DeepSeek V4.1 Flash 1017 remains first with 55.8 trillion prompt and completion tokens routed during the trailing week. Tencent’s Hy4 is second at 20.3 trillion, followed by the earlier DeepSeek V4.1 Flash at 16 trillion, Z.ai’s GLM 5.3 Flash at 12 trillion, and OpenAI’s GPT-6 Luna at 7.43 trillion. DeepSeek’s leading variant processed more tokens than the next three models combined.
That is a commanding position inside OpenRouter, not a universal market-share result. OpenRouter measures traffic through its own API, excludes private usage, and separates free and paid variants. Token volume is not users, revenue, requests, successful tasks, or preference. A coding agent reading a repository can burn orders of magnitude more tokens than a classification service handling thousands of short calls. The ranking is a map of routed workload, and maps become dangerous when somebody mistakes them for the territory.
The middle of the table is where the real experiment is running
The top-five freeze makes the movement below it more legible. Claude Opus 5.5 edged from seventh to sixth with 6.79 trillion tokens, passing Step 5 Preview at 6.73 trillion. The 60 billion-token gap sounds substantial until it is placed beside totals approaching seven trillion: it is less than one percent of either model’s weekly volume. This is a photo finish, not a migration story.
Xiaomi’s MiMo-V2.6-Flash offers stronger evidence of momentum. It climbed two places, from thirteenth to eleventh, while its trailing-week volume rose from 3.24 trillion to 3.62 trillion tokens. That 380 billion-token addition matters more than the ordinal change because it shows actual expansion rather than a model floating upward as competitors shrink. Even so, MiMo trails tenth-place GLM 5.3 by 260 billion tokens and leads MiniMax M3.1 Flash by only 20 billion. At this scale, eleventh place is a rented desk, not a corner office.
The counterexample is Crouton Alpha, the unidentified model that fell from eighth to thirteenth. Its rolling volume contracted from 4.5 trillion to 3.53 trillion tokens, a decline of 970 billion, while OpenRouter still labels it 125% above the preceding comparison week. Both statements can be true. The rank uses today’s trailing seven-day total; the percentage compares that rolling window with an earlier one. A large launch surge can remain far above the old baseline even while its newest daily traffic cools.
This is the first original lesson hiding in the table: rank direction and growth direction are different signals with different clocks. Teams building market intelligence from rolling dashboards should store daily snapshots, not scrape the displayed percentage and call it a trend. Without the underlying time series, an apparently contradictory “down five, up 125%” observation is easy to turn into a false narrative.
Hy4’s 405% is a decaying launch artifact, not a forecast
Tencent Hy4 remains second at 20.3 trillion tokens and shows 405% week-over-week growth. A day earlier, that displayed increase exceeded 999%. Hy4 has not suddenly stopped growing; its first heavy-usage days are moving into the comparison base. The denominator is catching up. Extrapolating 405% forward would be spreadsheet fan fiction.
The more durable signal is that Hy4 kept second place while the percentage normalized. Likewise, DeepSeek V4.1 Flash 1017’s rolling total rose from 53.2 trillion to 55.8 trillion even as its displayed comparison sat at negative one percent. Stable rank plus sustained absolute volume is more informative than one spectacular relative-growth badge.
Provider-level aggregation adds another trap. DeepSeek has three variants in the top 20 totaling 75.04 trillion tokens, but adding them does not produce a clean market share. Variants can differ by date, price, context window, routing policy, and workload. OpenRouter’s separate market-share view says DeepSeek handled 31.8% of text requests in the most recent complete week, followed by OpenAI at 19.3%, Google at 13.6%, Z.ai at 7.4%, and Anthropic at 5.8%. Request share and token share answer different questions. Long-context and verbose workloads inflate tokens without necessarily increasing the number of tasks served.
The second lesson is architectural: model popularity is partly a measurement of the applications wrapped around it. OpenRouter’s opted-in app table lists coding agents among its heaviest consumers—Claude Code at 1.35 trillion tokens, Kilo Code at 1.01 trillion, Codex at 694 billion, and OpenClaw at 387 billion. Agent loops read files, revise plans, call tools, retry failures, and carry growing context. A model that works well in those loops can dominate token volume without being the preferred choice for chat, retrieval, or low-latency structured extraction.
Use popularity to build a test queue, not a routing policy
For practitioners, the correct response is not to point production at this week’s highest climber. It is to use the table as a discovery mechanism. MiMo’s measured growth earns it a place in an evaluation queue. Crouton’s retreat is a reason to watch another week, especially while its provider remains undisclosed. Neither movement is a substitute for evidence from the workload that pays your bills.
A useful evaluation should start with a small regression set made from real tasks: representative repository changes, support conversations, extraction documents, or tool calls, including the ugly edge cases that caused incidents. Record cost per successful task rather than cost per token alone. Track p50 and p95 latency, time to first token, tool-call validity, retries, context-cache behavior, output length, and human correction rate. If the model is used agentically, cap the loop and measure how often it reaches the goal before exhausting that budget.
Keep model IDs configurable and preserve the exact endpoint in telemetry. Normalize those endpoints into model families for finance and reporting, but do not throw away variant detail. DeepSeek’s three ranked versions and Nvidia’s free Nemotron variants show why both levels are necessary. A dashboard that merges every family hides operational changes; one that treats every dated SKU as unrelated hides concentration risk. Record whether traffic is free, promotional, or paid, because a zero-price window can create usage that disappears the moment normal economics return.
The third lesson is about procurement rather than benchmarking. A marketplace leaderboard measures how readily developers can move work across suppliers. That makes the middle-table churn a useful proxy for substitutability. If a candidate can absorb real volume while teams switch through a common API, it may improve negotiating leverage even when it does not win your quality suite. The practical value of a credible second-choice model is often lower vendor risk, not a trophy for first place.
Arena AI’s leaderboard is unavailable today because its security verification blocked automated access, so there is no honest preference-ranking comparison to make. That missing panel should stay missing. Arena’s pairwise votes and OpenRouter’s traffic measure different things—what evaluators prefer versus what developers route—and recycling a stale Arena snapshot would manufacture precision rather than add context.
The week’s verdict is therefore less glamorous than “DeepSeek wins,” and more useful. The top of OpenRouter has consolidated; MiMo is gaining enough real volume to deserve testing; Crouton’s initial surge is cooling; and rolling percentages remain remarkably good at misleading anyone who ignores their windows. Popularity tells you where to look. Your own success rate, latency, cost, and failure data must decide where production traffic goes.
Sources: OpenRouter AI Model Rankings, OpenRouter Models API, OpenRouter API documentation, LM Arena leaderboard