The LLM Leaderboard Didn't Move. The Workloads Did.

A leaderboard can stay perfectly still while the market underneath it moves at sprint speed. That is the useful lesson in this week’s OpenRouter data: the five most-used models kept exactly the same positions, yet developers shifted enormous volumes of work toward a handful of challengers. If you only read the rank column, you miss the migration.

DeepSeek V4.1 Flash remains comfortably first with 36.8 trillion tokens routed over the trailing week, despite a 2% decline. Step 5 Preview is still second at 17.1 trillion—less than half DeepSeek’s volume—but its traffic grew 430% week over week. GLM 5.3 Flash, MiMo-V2.6-Flash, and Tencent’s Hy4 preview complete an unchanged top five at 12.7T, 10.9T, and 7.85T tokens respectively.

That is not a quiet week. It is a reminder that rank is a lagging, compressed representation of adoption. A model can add several times its previous workload and remain second because the incumbent’s lead is huge. Conversely, GPT-6 Luna rose from seventh to sixth while its traffic fell 6%, passing Nvidia’s Nemotron 3 Ultra free tier as that model slipped 1%. Movement on the table is not necessarily momentum; sometimes it is merely two declining lines crossing.

Step 5’s surge needs a production test, not a victory lap

Step 5 Preview’s 430% expansion is the table’s strongest signal, but it is not proof that the model is 4.3 times better—or even that five times as many people used it. OpenRouter counts prompt and completion tokens routed through its platform. Launch timing, promotional pricing, long context windows, verbose reasoning traces, and repeated retries can all inflate token volume. Different providers also use different tokenizers, making cross-provider totals directionally useful rather than laboratory-grade.

The model is still worth evaluating. Its OpenRouter endpoint supports reasoning controls, tool use, and structured response formats, the features that matter when a model graduates from chat window to agent workflow. Practitioner reaction has been appropriately mixed. Hacker News commenters welcomed the competitive pressure from open weights while noting that a 600-billion-parameter model with 27 billion active parameters is still a difficult fit for a 228GB local machine. OpenCode users have compared its capability with the GLM 5.3 Flash and DeepSeek V4.1 Flash tier, but complaints about latency and its economics without promotional pricing remain.

Those disagreements are not noise; they describe the evaluation you should run. Put Step 5 through representative agent traces, not a handful of prompts. Measure successful tasks per dollar, wall-clock time, tokens per accepted change, tool-call accuracy, retry rate, and how often a human has to repair the result. A model that produces a correct patch in one slow pass can beat a faster model that needs three attempts. A model that narrates every obvious step can lead a token chart while losing badly on production economics.

There is also a hard operational footnote: Step 5 Preview is not currently available through OpenRouter’s EU in-region routing. For teams with residency requirements, that fact can disqualify the model before quality testing begins. Compliance constraints belong at the top of a routing policy, alongside privacy and data-retention rules—not as an exception bolted on after a leaderboard winner has been selected.

Anthropic’s three-tier climb looks more durable

Step 5’s spike is concentrated in one model. Anthropic’s growth is spread across its current family: Claude Opus 5.5 sits eighth at 5.95T tokens and grew 38%; Sonnet 5.5 is fifteenth at 2.8T and grew 64%; Haiku 5.5 is eighteenth at 2.01T and grew 205%. When flagship, balanced, and fast tiers rise together, the plausible story is not merely launch curiosity. Developers may be sorting work by difficulty instead of sending every request to one default.

That pattern is more useful than copying any specific vendor choice. Route architecture, difficult planning, and final review to the strongest model; send routine implementation to the balanced tier; reserve extraction, classification, and lightweight transformations for the fast tier. Then evaluate the system as a whole. Cheap input tokens do not help if weak outputs create retries, while an expensive flagship is wasteful when a smaller model reliably handles the task.

The same principle explains why the day’s three small rank swaps are mostly trivia. Tencent’s Hy3 moved from twelfth to eleventh over Jev 1.13, but both display 3.25T tokens, so rounded totals conceal the actual margin. Muse Spark 1.3 Contributor moved from fourteenth to thirteenth over DeepSeek V4 Flash 0423. No model entered or left the top 20. The larger changes happened inside fixed positions: GLM 5.3 rose 53%, Hy3 54%, and Muse Spark 75%, while Claude’s family accelerated across all three tiers.

This is a demand sensor, not a quality benchmark

OpenRouter’s ranking measures public token traffic through one routing platform. It excludes private models, private endpoints, and zero-data-retention traffic; model variants are listed separately. It does not measure correctness, user count, revenue, total market share, or cost per completed task. The platform’s author-level request share—DeepSeek at 24.5%, OpenAI at 18.6%, Google at 17.8%, Z.ai at 13.4%, and Anthropic at 3.9%—uses requests rather than tokens, so those percentages should not be mixed with the token table as if they shared a denominator.

In other words, this leaderboard is closer to a demand sensor than a benchmark. It shows where a meaningful slice of public API experimentation and deployment is flowing, and the rate of change can identify models worth testing. It cannot tell you whether those models will solve your tickets, respect your latency budget, or behave consistently with your tools. Arena AI’s text leaderboard could not be independently verified for this edition—the dedicated collection returned no rows and a generic fallback timed out—so no stale Arena positions are being smuggled in to manufacture a broader consensus.

For engineering teams, the practical response is architectural. Keep model identifiers in configuration rather than application logic. Store per-task telemetry that includes provider, model version, latency, input and output tokens, retries, task outcome, and human correction time. Maintain explicit fallbacks, but test them under failure: a fallback that lacks a required tool schema or changes structured-output behavior is not a fallback, just a different incident.

Snapshots also need timestamps. OpenRouter updates continuously, returns large token totals as strings through its Data API to preserve 64-bit values, and can restate figures as traffic arrives. Save the meta.as_of value with every pull. Otherwise a dashboard that appears reproducible may quietly compare two different trailing windows.

The editorial verdict is simple: do not promote a model because it moved up one row, and do not ignore it because its row stayed put. Step 5 Preview’s growth deserves a serious bake-off; Anthropic’s broad three-tier rise deserves attention as a routing pattern; DeepSeek’s lead deserves respect without being mistaken for universal superiority. The winning teams will not hard-code this week’s number one. They will build evaluation and routing systems that can exploit next week’s change without turning it into a migration project.

Sources: OpenRouter Rankings, OpenRouter Data API documentation, Step 5 Preview on OpenRouter, Anthropic: Introducing Claude Opus 5.5, Hacker News discussion