Free Models Are Eating the Usage Chart — and That Is Both Signal and Noise
A model nobody can name just routed two trillion tokens in a week. That sounds like an intelligence breakthrough only if you confuse a distribution chart with a benchmark.
OpenRouter's latest seven-day rankings put the anonymous Ox Alpha at number nine immediately after its August 20 debut. The free-preview model arrived with the sort of specification sheet engineered to attract experiments: text, image, and video input; a 1,048,576-token context window; and as many as 131,072 output tokens. It also arrived with no price and no public provider identity. Mystery plus a zero-dollar meter is a remarkably effective developer-acquisition campaign.
Ox Alpha is not the only evidence. NVIDIA's free Nemotron 3 Ultra variant moved from seventh to sixth after processing 4.08 trillion tokens, up 111% week over week. Google's Gemini 3.7 Flash grew 975% to 1.49 trillion tokens and still slipped from thirteenth to fourteenth, while Gemini 3.6 Flash fell three places to thirteenth. The chart is moving quickly, but the movement says at least as much about pricing, promotions, migrations, and availability as it does about quality.
Two trillion tokens can hide a surprisingly small number of decisions
OpenRouter ranks prompt and completion tokens, not requests, users, revenue, or successful tasks. That distinction matters more than the ordering. A single agent run that reads a repository and emits a long patch can consume vastly more tokens than dozens of short classification calls. Tokenizers differ, models vary in verbosity, and a free endpoint removes the natural pressure to cap output. Private requests are excluded altogether. The result is a useful view of activity on one routing platform, but not market share and certainly not a referendum on which model is "best."
Ox Alpha's debut illustrates the denominator problem. At its advertised one-million-token context limit, two trillion tokens could theoretically represent roughly two million context-sized interactions. At a more ordinary 20,000 tokens per end-to-end agent run, it could represent about 100 million runs. Reality will be a messy mixture, but the range shows why volume without request count, unique users, or workload distribution cannot tell us whether broad adoption or a smaller cohort of token-hungry experimenters drove the spike.
That does not make the chart useless. It makes it a discovery feed. Models do not reach trillions of routed tokens without developers exercising them, integrations sending real workloads, or promotions creating enough curiosity to generate meaningful operational data. The mistake is treating discovery as evaluation. GitHub stars can tell you what to inspect; they cannot tell you whether a dependency belongs in your production graph. OpenRouter volume works the same way.
Free access is part of the benchmark now
Nemotron 3 Ultra shows why vendors can rationally buy distribution with inference. NVIDIA describes it as a 550-billion-parameter mixture-of-experts model with 55 billion active parameters, combining Mamba and attention, supporting a one-million-token context, using NVFP4 pretraining, and including native speculative-decoding MTP layers. It ships with pretrained, post-trained, and quantized checkpoints as well as training datasets. This is not merely a hosted black box dressed up as open infrastructure.
NVIDIA also claims Ultra delivers 5.9 times the inference throughput of GLM-5.1-754B-A40B, 4.8 times that of Kimi-K2.6-1T-A32B, and 1.6 times that of Qwen-3.5-397B-17B at an 8,000-token input and 64,000-token output. Those are vendor benchmarks and deserve independent verification. Still, they explain the economics: architectural efficiency can turn a free endpoint from pure subsidy into a relatively affordable way to collect developer feedback, expose compatibility bugs, and build routing mindshare.
This produces an uncomfortable but important metric collision. Operational efficiency lets a vendor give away more output. More output raises its token rank. The rank then looks like demand, even though part of the apparent demand is a consequence of supply-side economics. Promotional spend is not fake adoption, but it is not preference under normal prices either.
Google's transition from Gemini 3.6 Flash to 3.7 Flash adds a second confounder: forced or encouraged migration. The newer model carries introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens through year-end, half the original 3.6 Flash price. Google reports FrontierCode 1.1 Main improving from 34.4% to 43.6%, DeepSWE v1.1 from 49.0% to 65.3%, WebDev Arena Elo from 1538 to 1588, GDP.pdf from 22.0% to 34.0%, and AutomationBench from 17.0% to 30.4%. Those gains make a canary worthwhile. They do not make the price universally cheap: Ars Technica notes GPT-5.6 Luna at $0.20 per million input tokens and $1.20 per million output tokens.
The rankings capture that migration in progress. Gemini 3.7 Flash nearly grew tenfold while its predecessor suffered the week's largest incumbent decline. Looking at 3.7 alone understates Google's footprint before the switch and overstates net-new demand after it. Teams tracking the market should aggregate model families during replacement cycles, then separate versions again when evaluating regressions. Otherwise, a version rollout becomes a fictional market-share event.
What to test before the promotion becomes an architecture
For engineering teams, this week's movers should create an evaluation queue, not a procurement decision. Replay 50 to 200 representative tasks against the incumbent and candidate models. Measure task success, latency percentiles, retries, tool-call correctness, output length, and total cost. Keep failure categories specific: a syntactically valid tool call with the wrong argument is not a success, and a correct answer that consumes four times the output budget is not equivalent on cost.
Long-context claims need their own tests. Do not fill a million-token window with arbitrary text and celebrate when the request completes. Place relevant evidence at different depths, introduce plausible distractors, require citations back to the supplied material, and check whether retrieval quality collapses as the context grows. A large accepted payload is a capacity claim; reliable use of information inside it is a capability claim.
Ox Alpha warrants extra caution. Its frontier-shaped limits and early volume justify experimentation, but an anonymous preview should not become a production dependency until the provider discloses identity, data handling, retention, post-preview pricing, rate limits, and version-stability policy. Never send sensitive code or customer data merely because the endpoint is fashionable and free. Put it behind a configurable router, pin versions where possible, and maintain a tested fallback.
Gemini 3.7 Flash presents the inverse risk: a known vendor and concrete benchmark improvements, paired with a release cadence fast enough to create integration churn. The practical response is the same. Keep model choice in configuration, log the exact model returned, preserve a regression corpus, and canary changes before broad rollout. Three weeks between generations is impressive product velocity and a reminder that hard-coded model assumptions now expire faster than many sprint plans.
The metric worth watching next is retention after the incentives change. Does Ox Alpha remain in the top 20 after it gets a name and a price? Does Nemotron hold volume when free capacity tightens? Does combined Gemini 3.6 and 3.7 traffic grow once migration settles? One week measures acquisition. Four stable weeks begin to suggest product fit.
Usage rankings are valuable because they reveal what developers are willing to try at scale. They become misleading only when we ask them to answer a different question. This week's chart did not crown the smartest model. It showed that price, throughput, novelty, and migration are now first-class parts of model distribution—and that engineers need to separate the metrics that find candidates from the tests that decide what ships.
Sources: OpenRouter Rankings, Ox Alpha model page, Google's Gemini 3.7 Flash announcement, NVIDIA Nemotron 3 Ultra technical overview, Ars Technica