A Free Stealth Model Just Processed 6.54 Trillion Tokens. That Is Traction, Not Validation
A leaderboard can tell you where the traffic went. It cannot tell you whether the traffic arrived because a model was better, cheaper, newly free, or simply interesting enough for developers to point a few giant repositories at it. OpenRouter’s latest weekly ranking is a useful demonstration of that distinction: the anonymous, free Ox Alpha model jumped five places to number four after processing 6.54 trillion tokens, while Xiaomi’s MiMo-V2.5 rose to second with 9.14 trillion. Those are serious adoption signals. They are not quality scores.
The temptation is to translate the table into a horse race—Ox Alpha “beat” GPT-5.6 Luna, MiMo “passed” Tencent’s Hy3—but OpenRouter ranks models by prompt and completion tokens processed across the trailing seven completed UTC days. It does not rank by completed tasks, unique users, correctness, revenue, latency, or cost. Private traffic is excluded. Providers also use different tokenizers, which makes cross-provider totals approximate before we even ask whether a verbose model is doing more work or merely producing more text.
Still, the movement is too large to shrug off. Ox Alpha climbed from ninth to fourth as its displayed volume expanded from 2.0 trillion to 6.54 trillion tokens. That increase of roughly 4.54 trillion tokens pushed GPT-5.6 Luna down two places to sixth even though Luna’s volume held at 5.17 trillion. No model entered or left the top 20; Ox Alpha simply rearranged the upper half of the chart.
Free inference turns curiosity into trillions of tokens
Ox Alpha has nearly every ingredient required for a launch-window spike. OpenRouter describes it as a free reasoning model aimed at coding, sustained agentic work, and production workloads. It advertises a 1,048,576-token context window and a maximum output of 131,072 tokens. The provider remains identified only as “stealth.” Give developers a mysterious new model, a million-token context window, and a zero-dollar marginal API price, and they will test it on the largest inputs they can find.
That makes 6.54 trillion tokens meaningful, but not in the way a benchmark score is meaningful. Repository-scale coding agents repeatedly resend system prompts, source files, tool histories, and intermediate state. One long agent run can consume more tokens than hundreds of ordinary chat sessions. Free access also removes the natural governor that makes teams stop an unproductive loop. The ranking therefore measures consumption, and free agentic inference is very good at producing consumption.
Early community discussion reflects the right mix of curiosity and suspicion. Threads across LocalLLaMA, OpenAI, OpenCode, and OpenRouter communities are trying to identify the maker, with guesses ranging from GLM and MiMo to LongCat and other unreleased systems. None of those claims is verified. Comparisons against GLM 5.3, DeepSeek V4 Pro, Gemini 3.7 Flash, MiMo-V2.5-Pro, Qwen 3.8, and Kimi K3 are already circulating, but commenters have also noted that an approximately 15,000-token OpenCode system prompt can materially shape observed behavior. That is the less glamorous lesson: agent scaffolding is part of the benchmark, whether the benchmark author acknowledges it or not.
For engineering teams, the anonymity matters more than the rank. A stealth model may be fine for public-code experiments and synthetic tasks. It should not receive proprietary repositories, customer records, credentials, incident data, or unpublished product plans until its operator states who owns the service, how prompts are retained, whether data is used for training, where inference runs, and what support exists when something goes wrong. Free is a pricing decision, not a security review.
MiMo’s number-two position has a stronger durability case
MiMo-V2.5’s move from third to second looks less like pure launch tourism. Its weekly volume reached 9.14 trillion tokens, up from 8.18 trillion in the previous snapshot, and its displayed week-over-week growth accelerated to 138%. That was enough to pass Tencent’s Hy3 at 8.56 trillion, despite Hy3 itself growing 18%. DeepSeek V4 Flash 0731 remains first with 11.6 trillion tokens and 5% growth.
Xiaomi is positioning MiMo-V2.5 as a native multimodal agent model, with MiMo-V2.5-Pro aimed at complex coding and long-running work. Both support a one-million-token context and are released under the MIT license, permitting commercial inference, fine-tuning, and secondary training without additional authorization. Xiaomi also names day-zero support for SGLang and vLLM, plus hardware work spanning AWS Trainium2, AMD ROCm, and several Chinese accelerator vendors. That is not merely a model launch; it is an attempt to make hosted experimentation and later self-managed deployment part of the same adoption path.
The vendor’s long-horizon examples show what it wants buyers to notice. In one demonstration, the model built a SysY compiler in Rust over 4.3 hours, made 672 tool calls, and passed 233 out of 233 hidden tests. A separate video-editor task ran for 11.5 hours, invoked tools 1,868 times, and produced 8,192 lines. These are vendor-run demonstrations rather than independent benchmarks, so the correct response is replication, not applause. But they are more informative than a generic “coding” label because they expose the target workload: persistent state, repeated tool use, failure recovery, and enough runtime for small error rates to compound.
There is also a revealing flaw in the ranking itself. Xiaomi claims MiMo-V2.5 consumes about 50% fewer tokens than Muse Spark at the same ClawEval score, while the Pro model reportedly saves 42% versus Kimi K2.6. If those claims survive independent testing, a raw-token leaderboard systematically understates the value of the more efficient model. A model that finishes ten tasks with half the tokens may rank below one that finishes seven tasks while talking twice as much. Token volume is a billable resource, not a unit of useful work.
Build a workload leaderboard, not a fandom
Teams evaluating either model should use OpenRouter’s chart as a candidate-discovery feed, then move immediately to a fixed internal harness. Select representative repositories and real tool schemas. Include dependency failures, ambiguous requirements, permission errors, truncated outputs, and tasks that require the model to notice it is stuck. Measure completed tasks, hidden-test pass rate, human interventions, wall-clock time, tool-call validity, retries, cached and uncached input cost, output cost, and tokens per successful task.
Run the same agent scaffold across every candidate. Pin temperature, context-management rules, tool descriptions, retry policy, and maximum budget. Otherwise the evaluation mostly compares orchestration choices. For long-context models, test both a full-repository prompt and a retrieval-based setup; a million-token window is valuable only when the model can find and correctly use the relevant 200 lines. Large context can also make failures expensive, slow, and difficult to inspect.
The next several seven-day windows will clarify the story. Ox Alpha needs to remain near the top after its first full launch cohort rolls out, and its operator needs to identify itself. MiMo needs to retain usage once promotional incentives and novelty normalize. For both, task-specific adoption in code generation, debugging, shell execution, and multi-step planning would be stronger evidence than aggregate tokens. Watch cost per accepted task, not merely share of the token firehose.
A quieter signal sits on Arena AI, whose overall top 20 did not change in the latest snapshot. The rendered table did not expose trustworthy Elo values, vote totals, or a separable code leaderboard, so inventing precision would be worse than reporting none. Stability there alongside rapid OpenRouter reshuffling is a useful reminder that preference rankings and usage rankings answer different questions. One asks what evaluators choose in controlled comparisons; the other asks where public API traffic flowed under real pricing and launch conditions.
Ox Alpha’s 6.54 trillion tokens amount to an enormous public experiment, and MiMo-V2.5’s climb makes it a credible candidate for serious evaluation. Neither result licenses a production migration on its own. Benchmark them against your work, constrain their access, calculate success per dollar, and demand a named operator before trusting a stealth model with anything that matters. The market keeps mistaking consumption for quality. Engineers should know better.
Sources: OpenRouter LLM Rankings, Ox Alpha model page, OpenRouter Data API methodology, Xiaomi MiMo-V2.5 open-source announcement, MiMo-V2.5 technical examples, and Arena AI leaderboard. OpenRouter rankings data through August 22, 2026; licensed under CC BY 4.0.