Free Models Are Winning OpenRouter Traffic. That Doesn’t Make Them Better.
The most useful thing about this week’s model rankings is that they disagree.
Anthropic occupies seven of the top 20 positions on Arena’s visible overall leaderboard, including five of the first seven. OpenRouter’s usage chart tells a different story: its top five contains two DeepSeek models plus offerings from Tencent, OpenAI, and Xiaomi. NVIDIA’s free Nemotron variants are climbing, while a newly released Gemini Flash model has already entered the top 20.
That is not evidence that one leaderboard is broken. It is evidence that “best model” has become an underspecified question. Arena measures which answer people prefer in head-to-head comparisons. OpenRouter ranks the prompt and completion tokens processed through its API. The first is a signal about perceived output quality; the second is a record of what developers route when price, latency, availability, context length, and habit enter the room.
This week, economics entered loudly.
Free endpoints are buying something more valuable than attention
NVIDIA’s Nemotron 3 Ultra free endpoint rose two places to No. 8 on OpenRouter, processing 2.69 trillion tokens for the week and growing 20%. It edged Claude Opus 5, at 2.68 trillion, by only 10 billion tokens. At this scale that is a near tie, not a decisive overtake: one large batch workload or a busy day could reverse the order.
Farther down, Nemotron 3.5 Lightning free climbed from No. 20 to No. 18 with 836 billion tokens and reported growth above 999%. The percentage makes an excellent chart label and a poor planning input. OpenRouter warns that trend percentages can be distorted by a small comparison base, so the more important number is the absolute traffic. Hundreds of billions of tokens means developers are doing more than clicking through a launch demo.
Free access is often described as marketing, but for model providers it is also integration acquisition. Every prototype built around an endpoint creates prompts, evaluation fixtures, routing logic, and operational familiarity. Those are switching costs even when no invoice exists yet. NVIDIA is effectively paying the early adoption tax so developers can discover where Nemotron is good enough.
That does not make a free endpoint free infrastructure. Rate limits can tighten, pricing can appear, availability can change, and data-handling requirements may rule out third-party routing entirely. Teams evaluating these models should separate the experiment from the dependency: put the provider behind an internal interface, cap spend before a promotion ends, and retain a tested fallback. A temporary discount should not become a permanent architecture decision by accident.
Token volume rewards workload shape, not just model quality
DeepSeek V4 Flash 0731 remains OpenRouter’s No. 1 model at 11.3 trillion weekly tokens, ahead of Tencent’s Hy3 at 9.83 trillion. The 1.47 trillion-token gap is larger than the entire weekly volume of every model below No. 12. That dominance is real on this platform, but its meaning is narrower than the ranking invites readers to assume.
OpenRouter counts prompt and completion tokens in daily UTC buckets and excludes traffic that users or applications mark private. It does not count requests, unique users, spend, accuracy, or traffic sent directly to providers. Tokenizers differ. So does verbosity. A model used for long-context document processing can rank above one handling millions of short classification requests. A verbose model can generate more chart credit for delivering the same user value.
This creates a useful but underappreciated distinction: the chart measures consumed compute-shaped work, not product adoption in the conventional sense. One large synthetic-data pipeline may contribute more tokens than thousands of interactive users. Conversely, a compact, efficient model can be widely deployed and look smaller because it completes the job with fewer tokens.
For practitioners, OpenRouter’s table should be a discovery feed, not a selection algorithm. When a model climbs, add it to a controlled bake-off. Evaluate it on your actual request distribution and measure task success, quality per dollar, p50 and p95 latency, structured-output validity, tool-call success, safety failures, and retry rate. Token popularity can nominate a candidate. It cannot approve the pull request.
Version churn is now part of the operating model
DeepSeek V4 Pro 0813 rose three positions to No. 15 with 946 billion tokens. It sits only 3 billion tokens ahead of GPT-5.6 Terra, another photo finish dressed up as an ordinal ranking. Gemini 3.7 Flash debuted at No. 19 with 830 billion, while Gemini 3 Flash Preview fell three places to No. 20 at 821 billion. A 9 billion-token difference on an approximately 800 billion-token base is about one percent.
The interesting possibility is not that Gemini 3.7 Flash has “beaten” its predecessor. It is that users may already be migrating traffic between adjacent Google variants. Leaderboards increasingly capture release churn as much as stable demand. New versions arrive before teams have finished tuning prompts, collecting regressions, or understanding the failure modes of the previous one.
Constant upgrading feels modern and can quietly destroy experimental discipline. If the prompt changes with the model, the tool schema changes with the agent, and the evaluation set changes after a bad result, nobody knows what improved. Treat model identifiers like production dependencies: pin versions, keep prompts and evals in version control, shadow a small sample to new releases, and define promotion thresholds before viewing the outcome. Record cost and latency alongside quality so an apparent capability gain does not conceal an operational regression.
The same caution applies to the unchanged Arena order. Its visible top five remains Anthropic-heavy: claude-fable-5, two thinking variants of Claude Opus 4.6 and 4.7, Meta’s muse-spark-1.2 at xHigh, and Claude Opus 4.6. That is a meaningful preference signal, but not proof those models are the right default for extraction, code review, support classification, or an agent loop with strict latency limits. Human preference tests tend to reward polished answers; production systems often reward short, predictable, machine-checkable ones.
Use three gates, not one winner
A sensible model-selection process now has three stages. Use preference leaderboards such as Arena to find models with strong general capability. Use routing charts such as OpenRouter’s to spot ecosystem momentum, accessible endpoints, and models developers are testing at scale. Then make an internal evaluation the merge gate.
That internal gate should include failure cases, not just average scores. For a structured extraction workload, measure schema compliance and field-level accuracy. For coding, run repository-specific tasks and tests rather than generic puzzles. For agents, track tool-selection accuracy, loop count, recovery behavior, and total cost per completed task. Run the evaluation again when a provider changes a model alias or pricing tier. The goal is not to crown a universal champion; it is to identify the least expensive system that reliably clears your product’s quality bar.
This week’s movement is therefore less a horse race than a market diagram. Premium models continue to dominate preference tests, while cheap and free models attract enormous routed workloads. Providers are competing on intelligence, but also on distribution, migration friction, and the economics of experimentation.
The winners will not necessarily be the models with the highest leaderboard rank. They will be the ones teams can evaluate, afford, integrate, observe, and replace without drama. Read public rankings as signals about different parts of that system. Anyone compressing them into a single “best LLM” list has deleted the part engineers actually need to do.
Sources: OpenRouter AI Model Rankings, OpenRouter models catalog, Arena AI leaderboard