Trillion-Token Launches Are Outrunning the Model Leaderboards

The most useful model leaderboard this week is the one that did not move.

Arena AI’s overall top 20 remained unchanged, while OpenRouter’s trailing seven-day usage ranking recorded three swaps and two spectacular growth rates. Gemini 3.7 Flash climbed to number 13 on 1.35 trillion tokens, growing by more than 999% week over week. DeepSeek V4 Pro 0813 rose to number 15 with 1.08 trillion tokens and 559% growth. Neither move proves that these models are smarter than the ones they passed. It proves something more operationally relevant: developers can redirect enormous workloads toward a new model long before quality leaderboards reach a new consensus.

That distinction matters because leaderboards are increasingly being treated as interchangeable evidence. They are not. Arena measures human preferences in head-to-head response comparisons. OpenRouter ranks prompt and completion tokens routed through its marketplace during a trailing seven-day window, ending on the latest complete UTC day. One is a noisy proxy for perceived response quality; the other is a noisy proxy for adoption within one distribution channel. The disagreement between them is not a data problem. It is the story.

Trillion-token launches are becoming normal

OpenRouter’s top five did not change. DeepSeek V4 Flash 0731 remains first at 11.4 trillion weekly tokens, followed by Tencent’s Hy3 at 9.22 trillion and Xiaomi’s MiMo-V2.5 at 6.92 trillion. OpenAI’s GPT-5.6 Luna holds fourth at 5.6 trillion, and DeepSeek V4 Flash 0423 is fifth at 5.06 trillion. The first swap arrived at number nine, where DeepSeek V4 Pro 0423, with 2.11 trillion tokens and 24% growth, moved ahead of Gemini 3.6 Flash.

The smaller numbers lower down are more revealing. Gemini 3.7 Flash passed Kimi K3, while DeepSeek V4 Pro 0813 edged past Claude Sonnet 5. These are only one-position moves, but the velocity behind them is not subtle. Each new entrant converted launch attention into more than a trillion routed tokens within the measurement window.

The community attention was real, too. The main Hacker News discussion for Gemini 3.7 Flash reached 968 points and 495 comments; the DeepSeek V4 Pro 0813 thread reached 1,041 points and 453 comments. Those counts establish curiosity, not approval. Without sampling the discussion, it would be careless to convert engagement into a claim that developers love either model. Still, attention plus routed traffic tells us that trying a newly released model has become cheap enough—and integration layers flexible enough—for experimentation to happen at production scale.

Google’s position illustrates why “migration” is the wrong mental model. Gemini 3.7 Flash is growing at more than 999%, yet Gemini 3.6 Flash remains tenth with 2 trillion tokens and 18% growth. The new version has not simply displaced the old one. Usage appears additive, at least for now. Teams may be testing the revision, routing distinct workloads to each model, or retaining pinned versions while new deployments ramp. Model portfolios increasingly resemble overlapping release trains, not neat annual upgrades.

DeepSeek is building distribution, not just winning launches

DeepSeek now occupies four positions in OpenRouter’s top 15: first, fifth, ninth, and fifteenth. That is more significant than any single rank change. A provider with multiple Flash and Pro revisions receiving substantial traffic has portfolio-level distribution: users are not merely trying the newest flagship and leaving. They are selecting among cost, latency, and capability envelopes.

For engineering leaders, this creates an observability trap. Provider-level spend reports can make a system look stable while traffic shifts between dated model variants with different behavior. A move from one revision to another can change structured-output reliability, tool-call habits, latency tails, refusal patterns, or token consumption without changing the vendor name on a finance dashboard. Procurement can remain provider-level; production telemetry cannot.

It also complicates the popular idea of a permanent “default model.” When viable revisions arrive faster than a quarterly architecture review, a fixed default becomes a governance choice disguised as simplicity. That may still be the right choice for a regulated or regression-sensitive workload. But it should be explicit. Otherwise, teams either miss material price-performance improvements or absorb silent behavior changes through floating aliases and provider-side routing.

Popularity is a terrible router

OpenRouter itself warns that its ranking measures adoption, not accuracy, preference, request count, spend, or total market share. Private and zero-data-retention traffic is excluded. Token totals are also approximate across providers because tokenizers differ. A verbose model can generate more chart-visible volume than a concise one while completing the same number of tasks.

Free models make the caveat impossible to ignore. Nvidia’s Nemotron 3 Ultra free variant reached number seven with 3.75 trillion tokens and 94% growth. Nemotron 3.5 Lightning’s free entry grew 235% to 854 billion tokens. Both may be useful models. Both also benefit from a price that encourages broad experimentation, automatic fallback traffic, and workloads that would never be economical at paid rates. The chart cannot separate durable demand from a free sample table.

This is why copying the leaderboard into a routing policy is cargo cult architecture. Rank compresses several causes—quality, price, latency, availability, context length, launch promotion, and routing defaults—into one number. It does not tell you which cause applies to your workload. The same warning applies in reverse to Arena: a response preferred by anonymous voters may still be too slow, expensive, inconsistent at tool use, or difficult to operate under your data-retention requirements.

The right decision stack is straightforward. Use Arena as one external quality signal. Use OpenRouter as adoption telemetry. Let an internal task suite make the decision. If those three disagree, production traces win.

Treat model changes like dependency upgrades

Teams curious about Gemini 3.7 Flash or DeepSeek V4 Pro 0813 should run them now, but promotion should follow evidence rather than rank. Start with representative prompts and score task success per dollar, not subjective fluency alone. Measure p50 and p95 latency, structured-output validity, tool-call accuracy, retry behavior, and recovery after malformed tool results. Include difficult and boring cases; polished demos rarely fail where production systems do.

Pin the exact dated model identifier. Canary a small slice of real traffic, preserve the previous route, and make rollback a configuration change rather than a deployment project. Track output tokens as well as requests, because a cheaper per-token model can still cost more if it produces longer answers. Review provider fallback behavior and data-retention terms before sending production payloads. Then rerun the comparison over at least two trailing-week windows. A model that cannot hold usage after launch week generated marketing telemetry, not an architecture signal.

There is a broader engineering lesson in the stable Arena table and the moving OpenRouter chart. Model quality is becoming only one constraint in a multi-dimensional scheduling problem. Price, latency, reliability, privacy, regional availability, and operational familiarity can move real workloads even when preference scores stay put. The teams that benefit will not be those that chase every weekly winner. They will be the ones that build enough evaluation and routing discipline to test a fast mover without betting the system on it.

This week’s rankings do not crown a new best model. They show that adoption can move at trillion-token speed while quality consensus barely twitches. That gap is where serious model selection now happens—and it will not be resolved by staring harder at a single leaderboard.

Sources: OpenRouter rankings, OpenRouter Data API methodology, Arena AI leaderboard, Gemini 3.7 Flash discussion on Hacker News, DeepSeek V4 Pro 0813 discussion on Hacker News