The LLM Leaderboard Is Becoming a Routing Map

The most useful thing about this week’s model rankings is the disagreement inside them. OpenRouter’s traffic table says DeepSeek V4.1 Flash is the model moving the most work. Its separate intelligence index says Claude Opus 5.5 is the smartest model in the field. Both statements can be true, and teams that collapse them into a single “best model” list are making a procurement decision with the precision of a trending tab.

DeepSeek V4.1 Flash remains the clear traffic leader with 35.9 trillion tokens routed during the trailing week. That is 2.6 times the volume of second-place Step 5 Preview at 13.9 trillion, and nearly six times Claude Opus 5.5’s 6.05 trillion. The top five did not change: Step’s preview model is followed by GLM 5.3 Flash at 12.6 trillion, Xiaomi’s MiMo-V2.6-Flash at 10.3 trillion, and Tencent’s Hy4 preview at 7.55 trillion.

Yet the movement just below that group is more revealing than the static podium. Opus climbed from eighth to seventh as its traffic grew 29% week over week. It passed Nvidia’s free Nemotron 3 Ultra, which fell 19% to 5.34 trillion tokens. A paid frontier model taking share from a free model does not prove that developers suddenly stopped caring about price. It suggests that, for at least some workloads, the cost of a weak result is becoming more important than the cost of the inference that produced it.

A token leaderboard is a workload map, not an IQ test

OpenRouter is unusually clear about what its ranking measures: prompt and completion tokens, aggregated in UTC daily buckets, with model variants listed separately and private requests excluded. It does not measure unique users, revenue, successful tasks, latency, or correctness. A model can rise because one synthetic-data pipeline generated billions of tokens, because an agent got stuck in a verbose loop, or because a promotional free endpoint attracted experimentation. None of those scenarios necessarily means the model became better.

The request-share data makes the distinction harder to ignore. DeepSeek accounts for 24.5% of text requests in the most recent complete week shown, ahead of OpenAI at 18.6%, Google at 17.8%, and Z.ai at 13.4%. Anthropic sits at only 3.9% of requests even though three Claude models appear in the top 20. Request share and token volume describe different shapes of use: a small number of long, context-heavy agent sessions can generate more tokens than a large number of short classification calls.

There is another accounting trap. OpenRouter separates variants into individual rows. DeepSeek has three entries in the top 20, OpenAI has three, and Anthropic has three. Catalog structure therefore changes how concentrated a vendor looks. Add private enterprise traffic—which the public table cannot see—and this becomes a valuable view of one marketplace, not a census of the model economy.

That caveat does not make the data useless. It changes the question. The rankings are good at showing where developers are willing to route observable workloads. They are poor at telling you whether the resulting software worked.

Anthropic is building a three-speed adoption curve

Opus is only one part of the week’s Anthropic story. Sonnet 5.5 remains fifteenth with 2.69 trillion tokens, up 34%. Haiku 5.5 holds eighteenth with 2.02 trillion, but its 127% weekly growth is the fastest rate anywhere in the top 20. Together the three Claude models handled 10.76 trillion tokens—substantial growth, but still less than one-third of DeepSeek V4.1 Flash alone.

This looks less like a winner-takes-all race than a stack forming around task economics. Haiku can absorb cheap, frequent calls; Sonnet can serve the broad middle; Opus can take work where deeper reasoning, repository context, or recovery from a failed tool call justifies the premium. The important competitive advantage may not be owning one benchmark champion. It may be giving teams a relatively coherent ladder on which they can move workloads without redesigning every prompt and evaluation.

The community reaction around Opus also argues against a universal verdict. One practitioner described a coding workflow that produced 12 merge-ready pull requests, cut CI time from about ten minutes to four, and reduced billed CI minutes by roughly 60%. Other developers report the opposite experience: costly reasoning traces, overthinking, unreliable broad-context work, and plan limits that interrupt the workflow. These are anecdotes, not controlled trials, but their disagreement is useful. Model performance is increasingly a property of the model and the surrounding harness—context selection, tools, retry policy, tests, and the definition of done.

Meta’s Muse Spark 1.3 Contributor offers a smaller version of the same warning. It was the biggest rank mover, climbing two positions to eleventh on 3.2 trillion tokens and 58% growth. That is a meaningful demand signal. It is not enough evidence to replace a production model. No model entered or left the top 20 this week, so the broader picture is reallocation inside an established set, not a sudden market reset.

The right architecture is a routing policy

Engineering teams should stop asking which row wins and start measuring the cost of a completed task. Per-token price is an input, not the bill. A premium model that lands a correct patch in one attempt can be cheaper than a budget model that needs three retries, burns CI cycles, and consumes a reviewer’s afternoon. Conversely, using frontier reasoning to normalize addresses or classify support tickets is an expensive way to avoid writing a routing rule.

A practical evaluation should begin with real jobs sampled from production, not a generic benchmark. Split them into task classes such as extraction, summarization, code generation, repository-scale debugging, and tool-using agents. Run at least one high-volume value model and one frontier model against the same cases. Record task success, total input and output tokens, wall-clock latency, retries, tool failures, and human-review time. For coding work, include whether the patch passes tests and survives review; a model’s elo score will not tell you that.

Then route explicitly. Send high-volume, reversible work—classification, extraction, candidate generation, wide fan-out—to the cheaper model. Escalate when confidence is low, validation fails, or the expected cost of an error exceeds the frontier-model premium. Reserve the strongest model for ambiguous requirements, cross-file changes, incident diagnosis, and other cases where failure has operational consequences. Log the escalation reason so the policy can improve instead of becoming folklore.

Teams should also cap agent loops and track tokens per successful outcome. Token volume can hide pathological behavior: a verbose model may look heavily adopted while doing less useful work per token. Set maximum tool calls, wall time, and spend; require checkpoints before an agent expands scope; and compare models on identical context budgets. If a cheaper model needs twice the context and four retries, it may not be cheaper. If an expensive model spends thousands of reasoning tokens rediscovering information your retrieval layer should have supplied, that is a systems problem, not evidence of intelligence.

The week’s headline is therefore not that Claude beat Nemotron, or that DeepSeek remains number one. It is that model usage is settling into a barbell: enormous volumes flow through fast, economical models while premium reasoning earns a growing share of tasks where errors are expensive. The leaderboard can show that shift, but it cannot make the routing decision for you.

Buy intelligence where failure costs money. Buy throughput everywhere else. And judge both by completed work, because tokens are something a model spends—not something your users value.

Sources: OpenRouter LLM Rankings, OpenRouter model comparison, Hacker News launch discussion, Hacker News practitioner discussion