Step 5 Preview Did Not Climb OpenRouter's Chart. It Materialized at Number 12

Model leaderboards usually reward patience: a release arrives, developers test it, integrations follow, and usage climbs if the economics survive contact with production. Step 5 Preview skipped that sequence. In its first appearance on OpenRouter’s weekly chart, it landed at number 12 with 3.22 trillion tokens processed—more traffic than established releases from DeepSeek, Google, Moonshot, OpenAI, Meta, and Anthropic.

That is a remarkable launch. It is not, however, proof that Step 5 is the twelfth-best model in the world. OpenRouter’s ranking measures tokens routed through its platform, not answer quality, customer count, or successful tasks. A verbose model, a discounted launch, one high-volume application, or a default routing change can all manufacture a dramatic chart move. The useful signal is narrower and more practical: StepFun has found a price-and-capability envelope that developers were willing to put under real workloads almost immediately.

A trillion tokens can hide a surprisingly small number of decisions

The weekly window ending October 9 puts Step 5 Preview at 3.22T tokens, with 2.59T arriving on the latest complete day and daily usage rising 309%. The top of the chart remained stable—DeepSeek V4.1 Flash led with 37.6T tokens, followed by Space Bunny Alpha at 16.9T and GLM 5.3 Flash at 11.2T—but the lower half moved sharply. Claude Sonnet 5.5 debuted at number 20 with 1.71T tokens and 214% week-over-week growth. GPT-6.1 Sol rose from number 20 to number 18 as volume increased 311%, while GPT-5.6 Luna fell three places to number 17 after a 70% decline.

Those percentages look like a referendum. They are better read as migration telemetry. Router charts are unusually sensitive to concentrated demand: a handful of agent products can consume more tokens than thousands of interactive users because agents repeatedly read context, call tools, inspect results, and retry. That means Step 5’s debut may reveal less about broad developer preference than about one or two teams deciding that a previously uneconomic workflow is now affordable. For builders, that is still important. Production adoption often starts with a spreadsheet, a routing rule, and one application whose unit economics suddenly work.

StepFun’s offer is designed for exactly that calculation. The company documents a one-million-token input context, up to 64,000 output tokens, text, image, and video input, tool calling, JSON Schema output, prompt caching, and selectable low, medium, or high reasoning effort. Independent specifications describe a sparse mixture-of-experts system with 600 billion total parameters but 27 billion active for each token. API pricing is listed at $1 per million uncached input tokens, $0.05 for cached input, and $2.70 for output.

At those rates, a team can test repository-scale analysis, long research runs, and persistent agents without first negotiating flagship-model economics. Prompt caching is particularly consequential. Many agent systems resend the same repository map, policy bundle, or document corpus on every turn; reducing that repeated input to five cents per million cached tokens changes architecture, not merely the bill. It encourages longer-lived sessions and richer shared context.

There is a catch: context capacity is not context competence. A model accepting one million tokens says nothing about whether it can retrieve a buried constraint at token 480,000, preserve it after twelve tool calls, and apply it correctly in the final patch. Teams should test “needle” retrieval in the middle of realistic noisy inputs, not just at the beginning and end. They should also measure cache-hit rate, time to first token, tool-call recovery, and how performance degrades as the prompt grows. A giant window that requires duplicate reminders is expensive storage with a chat interface.

The benchmark table says “trial,” not “standardize”

StepFun reports 67.7% on DeepSWE v1.1, 33.3% on Terminal-Bench v4, 72.7% on SWE-Marathon v1.1, and a score of 1,571 on GDPval-AA v2. These are vendor-published results, and the comparison is not clean: Step 5 is run at High reasoning effort while some competitors use Max. The scores put it in credible company around models such as Kimi K3 and GLM 5.3 on several tests, but they do not establish leadership on difficult terminal and coding work.

Claude Sonnet 5.5 makes the trade-off visible. Anthropic prices it at $2 per million input tokens and $10 per million output tokens, says it generates more than 30% faster than Sonnet 5, and reports 70.6% on Terminal-Bench 4.0, up from 10.3% for its predecessor. The company also reports 46.2% on FrontierCode 1.1 at Max effort, 55.5% on CursorBench 4.0, and an 80.1% partial score on OSWorld 2.1. Slack’s Curtis Allen says internal Slackbot evaluations improved while using roughly 14% fewer output tokens and fewer steps.

This is why procurement by token price is a trap. Step 5’s output is roughly a quarter of Sonnet 5.5’s list price, but a model that needs four retries, emits five times as much text, or consumes twenty minutes of reviewer attention is not cheaper. Conversely, paying a premium for a stronger coding model is wasteful when the workload is document classification, evidence extraction, or a tool-driven report whose claims can be checked mechanically. The unit that matters is cost per accepted result: model charges plus retries, latency, infrastructure, and human review.

A sensible evaluation should use 30 to 100 representative tasks pulled from actual traces. Run each model against identical inputs and tool permissions, preserve the full transcript, and score completion rate, wall-clock time, input and output tokens, tool failures, retries, and reviewer minutes. Split results by task class. Repository navigation may favor one model; long-context synthesis or structured extraction may favor another. If the winner changes by workload, use routing rather than forcing one model to be the company religion.

“Open” is not useful until somebody can serve it

Step 5’s architecture also exposes a widening gap between available weights and deployable systems. In a Hacker News discussion, one practitioner’s reaction to the 600B-total model was blunt: “No way to run that on 228GB.” Sparse activation lowers compute per token, but it does not make hundreds of billions of stored parameters disappear. Quantization quality, interconnect requirements, expert routing, batching behavior, and serving software decide whether promised open weights become a practical option or an impressive download.

That matters for teams attracted by self-hosting, data control, or provider independence. Step 5 remains a preview, the independent hosting ecosystem is thin, and an independent audit found no linked technical report or named license at the time of review. Before sending proprietary repositories or customer records, engineers should verify retention terms, regional processing, provider logging, rate limits, and failover behavior. “Available through an API” and “safe to make a production dependency” are different milestones.

The next useful Step 5 headline will not be another rank. It will be reproducible evaluations from teams outside StepFun, weights released under a usable license, and independent hosts delivering the advertised quality and economics. Until then, the number-12 debut earns a benchmark slot, not an automatic migration.

OpenRouter’s chart is best understood as a map of where developers are spending inference, not a table of intelligence. This week it says cheap, cache-friendly, long-context agents can create enormous demand faster than the traditional benchmark cycle can explain it. Builders should follow that demand with discipline: test completed work, price the whole workflow, and keep an exit route. Trillions of tokens are evidence that somebody found a use case. They are not evidence that the use case worked.

Sources: OpenRouter rankings, StepFun model documentation, OpenRouter Step 5 Preview model page, CellCog specification and benchmark audit, Hacker News discussion, Anthropic’s Claude Sonnet 5.5 announcement