Free Tokens Won the LLM Ranking. That Doesn't Make Them Better.

A model leaderboard moved dramatically this week, but the most important number is not the rank. It is the price of admission: zero.

MiniMax M3.1 Flash jumped from tenth to fourth on OpenRouter after processing 10.3 trillion tokens in its first launch week. DeepSeek V4.1 Flash 1017, a dated snapshot released only days earlier, climbed to second with 16.5 trillion tokens. Those are real usage figures, and they show that developers are willing to test both models at enormous scale. They do not show that either model is the fourth- or second-best model in the world.

That distinction matters because OpenRouter's ranking measures prompt and completion tokens routed through its API. It does not measure active users, successful tasks, retained production deployments, latency, spend, or code that survived review. Different tokenizers and output styles change the tally. Private traffic is excluded. Free and paid variants are listed separately. The table is best understood as a map of distribution and experimentation, not a benchmark of intelligence.

MiniMax built a launch that was almost designed to win this chart

M3.1 Flash arrived on October 19 with a one-million-token context window and a free launch endpoint. Its trailing-week volume more than doubled in a day, from 4.97 trillion to 10.3 trillion tokens, producing the largest move in the top 20. OpenRouter still marks it as new because it does not yet have a complete prior-week comparison.

The combination is unusually favorable to a token-volume leaderboard. A free endpoint removes the cost brake, while a million-token context window invites users to submit entire repositories, documentation sets, and long agent traces. OpenRouter's activity breakdown makes that visible: 3.04 trillion prompt tokens, versus 19.4 billion reasoning tokens and 11.5 billion completion tokens in the displayed mix. The ranking records all that input as activity even if the resulting answer is abandoned after one attempt.

That is not an accusation of gaming. It is the expected result of aligning a generous launch offer with the metric being counted. The free period ended October 22, making the next week far more informative than the launch spike. M3.1 Flash now lists at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached-input tokens. If it remains in the top five after teams must attach a real budget to those giant prompts, its rank will start looking like durable adoption rather than subsidized curiosity.

The community response also argues against reading traffic as consensus. A Hacker News discussion reached 515 points and 219 comments, with several practitioners describing more initial mistakes, more coaxing, or weaker code-review and penetration-testing results than GLM 5.3 Flash and DeepSeek V4.1 Flash. The positive reports were specific rather than universal: one developer found parity with DeepSeek V4 Pro on an adversarial programming-language benchmark, while Plotly cofounder Chris Parmer reported that M3.1 Flash tied Opus 5.5 and Gemini Argon at 41 out of 43 on a data-analytics exam at roughly one-third the price. He also found it used many tokens and was the slowest model tested.

That last observation is the whole leaderboard problem in miniature. High token consumption can help a model climb the table while simultaneously making it slower or more expensive per completed job.

DeepSeek's split ranking looks more like migration than conquest

DeepSeek occupies the first two positions, but treating those rows as two independent adoption victories would be misleading. The base V4.1 Flash entry remains number one at 25.7 trillion weekly tokens, down from 27.3 trillion. The October 17 snapshot rose from 11.5 trillion to 16.5 trillion and moved from third to second. Because the regular deepseek-flash API alias now serves the newer build, some of the apparent growth is traffic migrating between labels inside a rolling seven-day window.

The release itself is operationally interesting. DeepSeek says the 1017 build improves benchmark performance, task-completion speed, and output brevity over its September version. It uses reinforcement learning with self-feedback for end-to-end project work and is optimized for DeepSeek Harness 0.2.1-rc.1 or later. The company also added Search and Fetch endpoints at $0.75 per 1,000 calls, while server-side tool calls through DeepSeek Harness, Claude Code, and Codex are free until October 31 at 16:00 UTC. The weights are available through Hugging Face and ModelScope.

For engineers, the important detail is the moving alias. A provider can call a release a Pareto improvement and still change output shape, tool-selection behavior, refusal boundaries, or assumptions embedded in prompts. Teams that need reproducibility should pin an explicit snapshot identifier, record the harness version, and promote new aliases only after replaying production failures. Model aliases are dependencies; treating them like an automatically safe package update is how an "improvement" reaches production before anyone measures the regression.

Rank movement is relative, and sometimes mostly mechanical

The rest of the table reinforces the point. Step 5 Preview fell nine positions, from eighth to seventeenth, alongside an 88% week-over-week decline. That drop lifted several models by one place without any corresponding surge in their own demand. Nemotron 3 Ultra's free variant climbed from twelfth to tenth even though its weekly volume fell 40%. It did not accelerate; competitors decelerated faster.

Meanwhile, GLM 5.3 Flash slipped from second to third with 11.3 trillion tokens, down 12%. MiMo-V2.6-Flash and Tencent's Hy4 preview each lost one position, despite Hy4 posting 9% weekly growth. Claude Opus 5.5 ranked seventh on 7.14 trillion tokens and 26% growth, while GPT-6 Luna sat eighth at 6.08 trillion and a 14% decline. These numbers are useful signals about what OpenRouter users are trying, but ordinal rank compresses very different stories—launch subsidies, snapshot migration, genuine growth, and relative decline—into the same arrow.

Arena AI offers a different lens through pairwise preference voting, but its data could not be verified for this edition: both the dedicated extraction path and a fallback encountered access failures. Repeating stale Elo scores would create a cleaner table and a worse article. More importantly, even a valid Arena score would answer a different question. Preference tests can estimate which response people favor under Arena's prompts; they still cannot tell a team whether a model will repair its codebase reliably inside its own tools and constraints.

Measure accepted work, not cheap tokens

Teams evaluating these models should use the ranking as a discovery queue, then build a workload-specific bake-off. Start with 30 to 50 representative tasks, including the failures and ugly edge cases that generic benchmarks omit. Freeze the harness version, temperature, tool permissions, timeout policy, and retry budget. Score correctness without revealing the model name, then add total cost and wall-clock latency.

For coding agents, track first-pass test success, intervention turns, tokens discarded in retries, p95 completion time, and cost per accepted or merged change. For support or extraction workloads, replace "merged change" with a domain outcome such as correctly resolved ticket or schema-valid record. A model priced cheaply per million tokens can be expensive if it rambles, retries, or consumes senior review time. A pricier model can be the budget option when its output passes on the first attempt.

There is also a sensible routing strategy hiding inside the noise. Use a cheaper flash model for routine, well-scoped tasks when your replay set proves it can handle them. Escalate planning, ambiguous debugging, and repeated failures to a stronger model. Cache stable context where the provider supports it, but do not use a giant context window as permission to skip retrieval design. Sending a repository because the meter is free is an experiment; sending it on every production call is an architecture.

MiniMax earned attention, and DeepSeek's new snapshot earned a serious test. Neither earned a procurement decision from this leaderboard. The signal worth watching now is persistence: whether MiniMax stays near the top after free access ends, whether DeepSeek's two rows consolidate as users migrate, and whether practitioners report fewer retries rather than merely more tokens. Until then, OpenRouter's chart tells us what developers are exercising. Your own traces must tell you what actually works.

Sources: OpenRouter LLM Rankings, DeepSeek V4.1 Flash 1017 release, OpenRouter MiniMax M3.1 Flash model page, Hacker News discussion