OpenRouter has launched Classifiers, a beta feature that sorts every request moving through its router into categories a workspace defines for itself: which team made the call, what kind of task it represents, or how complex the underlying agent behavior is. The result turns a raw activity log into something a finance or engineering lead can actually query: which team’s prompts are expensive, which model is absorbing which kind of work, and where spend is drifting.
A classifier combines four elements. A workspace sets how often traffic actually gets scored, since scoring every call is optional. Whichever model handles the scoring reads each captured request under instructions written specifically for that job, kept separate from whatever prompt generated the original response, and sorts it across as many as eight distinct dimensions rather than one flat tag. OpenRouter ships six ready-made presets, so a team does not have to design that structure first. Because scoring happens after the original response has already gone out, it does not slow the request down, and it keeps functioning even when prompt logging is switched off.
The underlying problem is not really about labels. Inference spend has become fragmented across models, providers, and internal teams in a way finance departments were not built to track. A company routing traffic through a single gateway to a dozen models can see total token spend easily. Seeing which product team, which internal agent, or which class of task generated that spend is a different, harder question. Most organizations currently answer it with workarounds: a separate API key per team so costs land on different invoices, a script that greps logs for keywords into a spreadsheet, or no attribution beyond a monthly total split by guesswork.
Classifiers replaces that guesswork with a model that reads the request itself and assigns categories, rather than requiring every caller to remember to pass a tag. That is a meaningful difference from how cost attribution usually gets built, since manual tagging schemes tend to decay as soon as a new team or agent starts calling the API without following the convention.
OpenRouter is not alone here. LLM gateway and observability products, including Portkey, Helicone, and self-hosted LiteLLM proxies, already let teams attach metadata to requests and build cost dashboards from it. Most of those tools require upfront discipline: the caller supplies the tag. OpenRouter’s bet is that automatic, model-driven classification survives the moments nobody remembers to tag anything, which is exactly when finance teams lose visibility.
For any team running production traffic through OpenRouter with more than one internal consumer, agent, or department, the next ninety days are a natural window to define a taxonomy and check whether the six presets actually match how the organization thinks about its own workloads before building a custom one from scratch.
OpenRouter announced Classifiers on X on July 27, 2026.
