Step 5 Hit No. 3 in One Week. That Is a Usage Signal, Not a Verdict

LLM leaderboards invite the wrong kind of certainty. A model moves from fourth to third, a rival slips one place, and the chart starts to look like a verdict. This week’s OpenRouter data says something more useful—and more limited: StepFun’s Step 5 Preview has attracted a remarkable amount of routed work, remarkably quickly. It does not say Step 5 is the third-best model in the world.

Step 5 Preview processed 12.1 trillion tokens in its first measured week, enough to take the No. 3 position from Xiaomi’s MiMo-V2.6-Flash at 11.5 trillion. Z.ai’s GLM 5.3 Flash remains second at 12.7 trillion, while DeepSeek V4.1 Flash is still operating in a different weight class: 40.5 trillion tokens, up 43% week over week. Step 5 has arrived at the front of the pack, but the leader is handling more than three times its volume.

The daily figures make the launch more interesting. On the latest complete day, Step 5 handled 3.25 trillion tokens, 32% more than the day before and well ahead of GLM 5.3 Flash’s 2.03 trillion. That was also more than half of DeepSeek’s 6.02 trillion daily total. Because the weekly table is a trailing window, a new entrant accelerating this quickly is partly hidden by its own incomplete history. If that run rate holds, the next weekly snapshot will tell a different story.

A usage chart measures workload, not wisdom

OpenRouter’s methodology counts prompt and completion tokens processed through its API. Private requests are excluded, model variants are counted separately, and the week-over-week percentage compares two trailing seven-day windows. It does not measure correctness, latency, reliability, unique users, or how many applications generated the traffic. One large batch-processing customer can outweigh thousands of developers trying a model once. A verbose model can also climb faster than a concise one while completing the same number of tasks.

That distinction matters because “No. 3” is easy to repeat and hard to unhear. The defensible conclusion is that Step 5 has won substantial distribution and experimentation on OpenRouter. It has earned a place in an evaluation suite. It has not earned an automatic production promotion.

The product shape helps explain the interest. OpenRouter describes Step 5 Preview as a sparse mixture-of-experts model with 600 billion total parameters but 27 billion active for a given token. It accepts text, images, and video, exposes a one-million-token context window, and supports up to 64,000 output tokens. It also covers the plumbing that determines whether a model can enter a real agent stack: tool use, structured outputs, log probabilities, stop sequences, and adjustable sampling.

Pricing is aggressive enough to encourage serious trials: $1 per million input tokens, $2.70 per million output tokens, and $0.05 per million cached input tokens. But Step 5 requires reasoning, with low, medium, and high effort settings. That turns the attractive sticker price into only one term in the cost equation. A cheap input token does not help if mandatory reasoning expands outputs, increases latency, or forces more retries. Production teams should measure dollars per successful task, not dollars per million tokens.

The million-token window is a test target, not a buying reason

Long context is especially vulnerable to marketing-by-maximum. Accepting one million tokens proves that an API will ingest a large payload; it does not prove the model will retrieve the right clause from page 4,000, preserve instructions across the prompt, or use tools correctly after a long conversation. Teams considering Step 5 for repositories, legal corpora, support histories, or multimodal documents should test retrieval at several depths and with distractors—not paste in a corpus, ask a friendly question, and declare victory.

A practical bake-off starts with 50 to 100 tasks sampled from actual traces. Freeze the system prompt and tool schemas. Score task completion with objective checks where possible, then record first-token latency, total latency, valid tool-call rate, retries, input and output tokens, and cost per successful result. Test low reasoning effort on routine classification and dispatch, then medium or high effort on planning and multi-step tool work. The model should earn extra latency task by task, not through a global setting chosen because “reasoning” sounds safer.

There is a second operational issue hidden inside the ranking: previews move. Step 5’s feature set makes it easy to slot into an OpenAI-compatible harness, but compatibility is not reproducibility. Pin explicit model identifiers where the provider allows it, retain captured evaluation inputs, enforce output budgets, and keep a provider-diverse fallback. A model that enters this quickly can change just as quickly, whether through routing, pricing, rate limits, or a revised preview.

The rest of the table is a warning against rank theater

Nvidia’s free Nemotron 3 Ultra moved above GPT-6 Luna into sixth place, 7.17 trillion tokens to 7.11 trillion. That 60-billion-token gap is only about 0.8% of their roughly seven-trillion totals. Calling one the winner would turn rounding-level movement into a horse race. Treat them as tied until repeated windows show separation.

Space Bunny Alpha provides the opposite lesson. One day after appearing at No. 10, its weekly volume fell 89% and it vanished from the top 20. Every model previously ranked from 11 through 20 rose one place except the departed model; GLM 5.2 entered at No. 20 with 1.53 trillion tokens. Most of that green-arrow movement is mechanical, not evidence that nineteen products simultaneously became more popular.

Meanwhile, the strongest percentage growth is below the podium. Claude Opus 5.5 rose 106% to 5.35 trillion tokens, Claude Sonnet 5.5 rose 186% to 2.65 trillion, and GPT-6.1 Sol rose 140% to 2.23 trillion. Percentage growth from a smaller base is not dominance, but it can reveal where new workloads are forming before absolute rank catches up. Engineers evaluating the market should watch sustained volume over several windows, not a single rank or growth badge.

The provider mix is also strategically useful. Four providers occupy the top five positions, while Anthropic and OpenAI are accelerating farther down the list. OpenRouter users are not the whole market—they likely overrepresent teams already comfortable with multi-provider routing—but the chart still shows that model distribution is becoming less synonymous with one vendor’s native API. The payoff for application teams is negotiating power. Once tool schemas, prompts, traces, and evaluations are normalized, changing models becomes a controlled engineering decision rather than a rewrite.

There is no verified Arena AI Elo table to pair with today’s usage data. The approved scraper returned no rows and the fallback was blocked by Cloudflare, so publishing remembered or stale numbers would manufacture precision. There was also no verifiable Hacker News, Reddit, or X thread with stable engagement counts in the available research path. The honest evidence is the traffic itself: substantial workload, unclear concentration, and no independent proof here of superior quality.

Step 5 Preview deserves attention because 12.1 trillion tokens in a first measured week is hard to dismiss, and a 3.25-trillion-token latest day suggests the weekly chart has not caught up. But the correct response is disciplined curiosity. Add it to the test matrix for tool-heavy, long-document, and multimodal work. Compare successful-task economics, stress the long context, preserve fallbacks, and check again after a full second week. Step 5 has earned a serious code review. It has not yet earned LGTM.

Sources: OpenRouter LLM Rankings, OpenRouter model catalog API, Step 5 Preview model page, OpenRouter rankings methodology