Baidu Bought Its Way Into the Usage Chart — Now We Find Out Whether Developers Stay

Free inference is an excellent way to win a usage leaderboard and a terrible way to prove that developers love your model. Baidu’s ERNIE-5.5-Flash-Preview entered OpenRouter’s weekly top 20 at number 20 after processing 2.47 trillion tokens. That is serious distribution, but it is not yet evidence of quality, retention, or willingness to pay.

The distinction matters because model charts are increasingly treated like sports standings: one number, one rank, one apparent winner. OpenRouter’s table measures prompt and completion tokens routed through its platform in UTC daily buckets. It excludes private requests and separates variants such as free editions. It does not measure accuracy, unique users, revenue, task completion, or satisfaction—and OpenRouter itself warns that token volume is not a quality score.

So the useful reading of ERNIE’s debut is narrower and more interesting than “Baidu has the twentieth-best model.” Baidu has removed enough friction for developers, routers, classrooms, demos, and high-volume pipelines to try the model at enormous scale. Distribution has cleared its first hurdle. The next question is whether the traffic survives contact with production.

Free models manufacture trial; production systems measure regret

Price changes behavior before capability does. A free endpoint can become the default for document extraction, synthetic data, test fixtures, hobby projects, and agent loops where a paid call would feel wasteful. Router policies also amplify the effect: when the marginal inference price is zero, a model can receive work it would never win in a head-to-head quality evaluation.

That does not make 2.47 trillion tokens fake. It makes them a measure of trial rather than commitment. ERNIE also ranks first on OpenRouter’s “New & Trending” view because it has no prior-week baseline. A new free SKU is mathematically designed to look explosive during its launch window. The durable signal will arrive later: whether it remains in the top 20 after a full comparison period, whether usage persists if pricing changes, and whether it appears in task-specific views rather than only aggregate traffic.

This is the first original lesson in the chart: token volume has a hidden denominator. A verbose model can consume more tokens to finish the same work. Long-context summarization and synthetic-data jobs can dominate volume while involving relatively few users. A model that processes twice the tokens may be serving more demand, producing longer answers, handling different workloads, or simply wasting context. Without completed tasks and user counts, the chart cannot tell those stories apart.

For engineering teams, the right response is not to ignore the ranking. It is to use it as a discovery feed. Take ERNIE’s entry as a reason to evaluate the model, then make the production decision with your own evidence. Build a replay set from real, sanitized workloads: tool calls, long-context prompts, malformed input, refusal edge cases, multilingual requests, and strict structured-output schemas. Measure validated completion rate, retry frequency, p50 and p95 latency, tokens per successful task, and fallback usage. “Free” becomes expensive quickly when a failed JSON response triggers another call, a human review, or an incorrect downstream action.

DeepSeek’s lead is less a horse race than a traffic pattern

The larger number in this week’s data belongs to DeepSeek V4.1 Flash 1017. It handled 61.2 trillion tokens, up 10% week over week. Tencent’s Hy4 remained second at 16 trillion, while the older DeepSeek V4.1 Flash sat third at 15.5 trillion after a 12% decline. The leading DeepSeek variant processed roughly 3.8 times the runner-up’s volume and about as much as the next five models combined.

The top ten did not reorder. OpenAI’s GPT-6 Luna grew 14% to 8.08 trillion tokens, and Anthropic’s Claude Opus 5.5 grew 9% to 7.37 trillion, but neither came close to changing the shape of the chart. Xiaomi’s MiMo-V2.6-Flash was the fastest-growing established top-ten model, rising 63% to 4.16 trillion tokens while holding eighth place.

Provider-level request share adds another angle. DeepSeek accounts for 31.8% of text requests by author on OpenRouter, ahead of OpenAI at 19.0% and Google at 15.1%. That figure counts requests rather than model-table tokens, so it should not be mixed casually with the 61.2-trillion figure. Still, both views point toward the same operational reality: inexpensive, fast models are absorbing an enormous amount of routine throughput.

But task-specific spending tells a different story. In OpenRouter’s implementation-workload view, Claude Opus 5.5 leads with 25.6% of spend, followed by Kimi K3.1 Pro at 7.9% and Claude Sonnet 5.5 at 6.7%. Aggregate popularity and coding preference are not the same market. Builders appear willing to route bulk tokens toward fast or economical models while paying premium rates for harder implementation work, where a failed attempt costs more than the inference bill.

That split suggests a second conclusion: the era of one global model default is ending. A production stack should route by failure cost and workload shape. Cheap models can handle classification, extraction, summarization, and low-risk generation; stronger models can take ambiguous code changes, tool-rich agents, and decisions with expensive consequences. The architecture matters more than whoever occupies number one this week. A modest router with clear escalation rules can outperform a supposedly superior universal default on both reliability and cost.

A rank can rise while demand collapses

The movement just outside the top ten is a useful warning against reading arrows without reading numbers. MiniMax M3.1 Flash climbed from number 12 to number 11 even though its volume fell 28% to 3.35 trillion tokens. It rose because Kimi K3.1 Pro fell faster—26% to 3.24 trillion—and slipped to number 12. Nvidia’s free Nemotron 3 Ultra moved from 14 to 13, while DeepSeek V4 Flash 0731 moved the other way.

A green arrow can therefore describe relative decay, not growth. Teams comparing providers should store the underlying weekly values, not merely scrape rank. Better still, calculate share of tracked volume and multi-week retention. Rank is ordinal: it hides whether two models are separated by a rounding error or an order of magnitude. DeepSeek’s gap over Hy4 and MiniMax’s narrow move past Kimi both produce a one-line leaderboard fact, but they describe entirely different competitive situations.

There is also a missing cross-check this week. Arena’s text leaderboard could not be verified: the dedicated data path returned no results and the fallback was blocked by Cloudflare. No preference ranks or Elo values should be inferred from old data. That absence is inconvenient, but publishing a blank is better than quietly carrying forward stale numbers. OpenRouter shows what developers routed through one marketplace; Arena, when available, provides a different signal based on comparative user preferences. Neither is a substitute for workload-specific evaluation.

Teams considering ERNIE should watch three checkpoints. First, does its volume hold after the launch window and any free-access changes? Second, does it earn traffic in coding, analysis, and tool-use categories rather than only undifferentiated token totals? Third, do independent evaluations corroborate the trial spike with acceptable reliability and preference results? Until then, the model is a credible candidate, not a proven default.

Baidu bought the cheapest and often hardest part of adoption: getting developers to try something new. The 2.47 trillion tokens show that the offer worked. What happens next will reveal whether the product earns repeat traffic once engineers count retries, latency, schema failures, and the real cost per completed task. Leaderboards are useful when they start an investigation. They become dangerous when the rank is mistaken for the conclusion.

Sources: OpenRouter AI Model Rankings, OpenRouter Models API, OpenRouter documentation, LM Arena leaderboard