Free Traffic Fell Away. Paid MiniMax and Step 5 Took the Ranks Instead

The most useful thing about this week’s model leaderboard is not which model “won.” It is that two versions of the same MiniMax model moved in opposite directions: the free endpoint fell six places while the paid endpoint climbed seven. That is less a verdict on intelligence than a clean view into how price, quotas, routing, and production reliability shape the models developers actually use.

OpenRouter’s weekly usage ranking kept its top five intact through October 28. DeepSeek V4.1 Flash 1017 remained first with 39.6 trillion tokens, up 140% week over week, followed by the older DeepSeek V4.1 Flash at 16.6 trillion. GLM 5.3 Flash, GPT-6 Luna, and Tencent’s Hy4 preview completed the unchanged top tier. The movement happened below them, where Step 5 Preview jumped from eleventh to seventh on 5.39 trillion tokens, a 132% increase, and paid MiniMax M3.1 Flash rose from twentieth to thirteenth on 3.21 trillion tokens and growth above 999%.

Meanwhile, MiniMax M3.1 Flash’s free variant dropped from sixth to twelfth after its weekly volume contracted 68%, landing at 3.33 trillion tokens. The free SKU still carried slightly more traffic than the paid one, but only by about 120 billion tokens—a rounding error at this scale. A week ago they occupied very different parts of the chart. Now their volumes are nearly equal and their trajectories point in opposite directions.

The chart is infrastructure telemetry, not an IQ test

OpenRouter is explicit about what its ranking measures: prompt and completion tokens routed through its public marketplace. It treats free and paid variants separately and excludes private activity. It does not measure unique users, revenue, accuracy, latency, successful tasks, or reasoning quality. That distinction matters because token volume is unusually easy to misread.

A coding agent working for 20 minutes can consume more tokens than hundreds of short chat requests. A verbose model can climb the chart while completing fewer useful jobs. A router default, free-tier quota change, or widely used application switching providers can move trillions of tokens without any underlying capability change. Mandatory reasoning can inflate usage further because the model spends tokens on intermediate work whether or not that work improves the result.

That makes the MiniMax split a supply-and-economics signal first. The paid M3.1 Flash listing offers a 1,048,576-token context window, text, image, and video input, and mandatory reasoning at $0.15 per million input tokens and $0.60 per million output tokens. Those prices are low enough that a team graduating from a promotional endpoint may accept the bill in exchange for steadier capacity. If a free route becomes quota-constrained, slower, or less available, production traffic does not need to abandon the model family; it can simply cross the narrow bridge to the paid SKU.

This is the first practical conclusion from the divergence: “free” and “paid” are not interchangeable deployment targets even when the model name is identical. They can have different rate limits, provider routes, queue behavior, availability, and operational guarantees. Benchmarking the free endpoint and deploying the paid one—or the reverse—leaves a hole in an evaluation that otherwise looks rigorous.

Step 5’s rise looks durable enough to investigate

Step 5 Preview’s four-place gain deserves attention for a different reason. Its 5.39 trillion weekly tokens were backed by 26.5 trillion over the trailing 30 days. That does not prove quality, but it makes the move harder to dismiss as a one-day experiment or a brief burst from a single launch campaign. Enough workloads are sticking around to put Step 5 ahead of established names including Claude Haiku 5.5 and Claude Sonnet 5.5 on this particular usage measure.

The right response is not to swap a production model because an arrow turned green. It is to add Step 5 Preview to a workload-specific bake-off. Take 20 to 50 real traces—support conversations, code changes, extraction jobs, or tool-using agent runs—and replay them against the exact model and route under consideration. Score task completion, tool-call validity, factual error rate, p50 and p95 latency, retries, and total tokens per accepted result.

That last metric is the one leaderboard watching tends to obscure. Cost per token is an input price, not an outcome price. A cheap model that emits twice as many tokens, fails a tool schema, and needs a repair pass can be more expensive than a premium model that completes the job once. For agent workloads, the useful denominator is dollars per successful task, including retry tokens, orchestration overhead, and the human time spent reviewing failures.

There is also a preview-risk tax. A preview model can be attractive for experiments while remaining a poor dependency for a product with strict regression requirements. Teams testing Step 5 should pin the exact identifier, store representative outputs, and be ready for behavior to change. If the provider does not guarantee a stable snapshot, the eval suite needs to run continuously rather than only during initial selection.

DeepSeek’s lead is becoming an availability concern

DeepSeek’s two leading variants processed 56.2 trillion weekly tokens combined, and the provider held four of the top 20 slots. That concentration is operationally meaningful even for teams that never select DeepSeek by name. Routers, coding tools, and third-party applications may depend on the same provider capacity, turning an outage or policy change into a shared failure mode.

Production systems should therefore keep model IDs configurable rather than burying them in application code, capture cost and quality telemetry per model route, and maintain at least one tested fallback from a different provider. “Fallback” must mean more than an alternate string in a configuration file: replay the same eval set, verify tool schemas, check context-window assumptions, and test the failover path before an incident. A backup that has never handled real traces is documentation, not resilience.

The contrast with benchmark rankings reinforces the point. OpenRouter’s separate Artificial Analysis index places Claude Opus 5.5 first at 58.4 and GPT-6 Astra second at 56.9, while the usage chart is dominated by open-weight and aggressively priced models. These lists are not contradicting each other. They answer different questions: one approximates capability under a benchmark methodology; the other records where tokens flowed under real marketplace constraints.

Arena AI’s text leaderboard could not be independently verified in today’s run because its automated endpoint returned no rows and the fallback hit Cloudflare verification. Leaving that table blank is the correct editorial choice. Stale or guessed Elo positions would add false precision to a story whose central lesson is already about reading measurements carefully.

For engineering teams, this week’s shortlist is straightforward: evaluate Step 5 Preview and paid MiniMax M3.1 Flash if cost-sensitive agent work matters; benchmark the exact paid or free route intended for deployment; measure successful-task economics rather than token price; alert on router substitutions; and keep a cross-provider fallback warm. None of that requires believing popularity equals quality.

The leaderboard is valuable precisely when it is not treated as a medal table. It shows where infrastructure, economics, and developer behavior are converging. This week, the sharpest signal is that “free” traffic can disappear while paid usage accelerates, and that a model can become operationally relevant before benchmark consensus catches up. Use the chart to decide what to test next—not what to trust blindly.

Sources: OpenRouter AI Model Rankings, OpenRouter model catalog API, MiniMax M3.1 Flash model page, OpenRouter data methodology