The Top Five Stayed Frozen. The Routing Layer Did Not.

The most interesting thing about this week’s model rankings is that almost nothing happened at the top. OpenRouter’s five most-used routes held their positions: DeepSeek V4.1 Flash 1017, DeepSeek V4.1 Flash, GLM 5.3 Flash, GPT-6 Luna, and Tencent’s Hy4 preview. That stability looks like consensus until you inspect the traffic underneath it. The leaderboard is less a verdict on model intelligence than a live map of production routing decisions—and this week that map shows migrations, not coronations.

DeepSeek V4.1 Flash 1017 remained No. 1 with 42.4 trillion tokens for the measured week, roughly 2.7 times the 15.9 trillion processed by the older DeepSeek V4.1 Flash route at No. 2. The newer route grew 96% week over week while the older one declined 32%. That is not broad-based growth for a model family. It looks much more like traffic consolidating around a refreshed endpoint.

This distinction matters because OpenRouter ranks routes and variants separately. Its weekly table counts prompt and completion tokens in daily UTC buckets and excludes private activity. A free route, a paid route, and a dated model revision can therefore move in opposite directions even when users think of them as “the same model.” The rows describe consumption through one routing platform, shaped by price, availability, application defaults, and rate limits. They do not measure reasoning quality.

MiniMax’s paid route inherited the workload

The clearest movement came from MiniMax M3.1 Flash. Its paid route jumped four places, from No. 13 to No. 9, on 4.12 trillion weekly tokens. OpenRouter displayed growth above 999%. Meanwhile, MiniMax M3.1 Flash (free), previously No. 12 with 3.33 trillion tokens, disappeared from the top 20 entirely.

It would be tempting to frame that as developers suddenly discovering MiniMax. The paired movement suggests something more operational: existing demand moved from one route to another. A change in free capacity, quotas, availability, or application routing can create a spectacular percentage without a corresponding change in user preference. The underlying lesson is blunt: free endpoints are acquisition channels, not service-level agreements.

Teams depending on a free route should check three things now. First, confirm that production traces record the exact route and version instead of only a family alias. Second, exercise the fallback chain under real quota and timeout conditions; a fallback that exists only in configuration is not resilience. Third, calculate the paid route’s cost per completed task, including retries. The meaningful unit is not dollars per million tokens if the cheaper route produces malformed tool calls and requires a second run.

MiniMax’s move also exposes a recurring weakness in usage leaderboards: they can count a migration twice in the narrative. One route falls, another rises, and observers describe both as independent shifts in taste. Engineers should instead ask whether the workloads, clients, and prompts changed. Without that context, a four-place jump may be a routing event wearing the clothes of product momentum.

GPT-6.1 Sol has acceleration, not yet a trend

GPT-6.1 Sol rose three positions to No. 12 with 3.75 trillion weekly tokens, up 76%. Its most recent complete day was stronger: 1.13 trillion tokens, up 114% day over day and enough for eighth place on the daily table. GPT-5.6 Luna also crossed into the weekly top 20 at No. 20 with 2.24 trillion tokens, up 36%. More than two trillion tokens were required merely to occupy the last visible row.

Sol’s daily acceleration makes it worth testing, but one week does not justify a production rewrite. The useful response is a shadow evaluation. Replay representative prompts, preserve the real system instructions and tool schemas, and score task completion rather than stylistic preference. Record p50 and p95 latency, tool-call validity, structured-output failures, retry frequency, and cost per successful job. If the model wins on those measures for a sustained period, then promote it gradually.

This is especially important because agent applications are now large enough to move the rankings themselves. OpenRouter’s top-app table showed Codex at 2.62 trillion tokens, Claude Code at 2.13 trillion, and OpenClaw fifth at 731 billion. A default-model change in one major coding agent can shift several leaderboard rows without millions of developers making an explicit choice. The leaderboard increasingly measures software defaults as much as human preferences.

That creates a feedback loop. A model becomes the default in a popular client, usage rises, the higher ranking attracts more evaluators, and some of those evaluators adopt it. None of this is illegitimate, but it means rank movement should be treated like package-download statistics: valuable evidence of ecosystem activity, weak evidence of fitness for your workload.

The declines are operational signals too

Xiaomi’s MiMo-V2.6-Flash fell three places to No. 13 as its weekly volume dropped 55% to 3.58 trillion tokens. Nvidia’s Nemotron 3 Ultra (free) slipped two places while volume declined 15%. Those decreases do not prove either model regressed. They can reflect expiring promotions, capacity pressure, a client changing defaults, or traffic moving to a newer sibling.

Still, falling usage can matter to production teams even when quality is unchanged. Lower aggregate demand may improve latency, but it can also precede reduced investment in a route, fewer integrations, or deprecation. If a declining model sits in a critical workflow, check provider notices, model lifecycle guarantees, and the availability of compatible replacements. Run the replacement through the same regression suite before urgency turns migration into an incident.

DeepSeek’s split is the most consequential version of this problem. The 1017 route processed more tokens than the next two entries combined, while the previous route shrank. Teams using a family-level alias should verify whether their provider silently advanced the backend. Rerun tests for JSON conformance, tool selection, refusal behavior, long-context retrieval, and token accounting. Compatibility labels are convenient, but a supposedly minor revision can still change the failure modes that matter.

Keep popularity, preference, and reliability on separate dashboards

Arena AI could not be verified for this report. Its dedicated scraper returned no data, and the fallback hit Cloudflare verification. That missing comparison matters: Arena is designed to capture human preference through pairwise evaluation, while OpenRouter reports routed usage. Claiming those signals moved together would be synthetic precision, so no Arena movement is asserted here.

The practical playbook is to maintain three distinct views. Use public routing data to spot models and routes gaining operational adoption. Use controlled evaluations to measure quality on your own tasks. Use production telemetry to measure reliability and economics. Combining those into one “best model” score produces a neat chart and an expensive decision.

Export your own seven-day routing mix and compare it with the public table. Add the exact provider, route, and model revision to every trace. Evaluate models that show sustained multi-week growth, not merely one hot day. Test failover while deliberately injecting quota errors and timeouts. Above all, treat route availability as part of the model contract: an excellent model behind an unstable endpoint is a poor production dependency.

The static top five do not show a market converging on one universal winner. DeepSeek, Z.ai, OpenAI, Tencent, and Anthropic all retain major traffic, while the lower ranks rotate as applications optimize for cost, speed, and availability. This is what a mature routing market looks like: less religious allegiance, more workload-specific substitution.

The verdict for this week is therefore not “MiniMax won” or “Sol is better.” It is that routing decisions have become powerful enough to masquerade as model breakthroughs. Read the table as evidence of where workloads are moving, then make your own model earn deployment with task-level data. Popularity is useful reconnaissance. It is not an acceptance test.

Sources: OpenRouter AI Model Rankings, OpenRouter Data API documentation, OpenRouter model catalog, LM Arena leaderboard