Published · 1h agoScience2 min read
No $7.5B in the docs: what OpenRouter's 7-day spend window does to your backend
The supplied OpenRouter documentation carries no price at all. What it does carry is a router that picks your model from the last seven days of other people's spending.
Written for builders.See today for builders
What happened
- The Auto Router is invoked by sending openrouter/auto as the model, and works like any other model slug; openrouter/auto-beta is the early-access track.
- A fast, lightweight classifier assigns each prompt one of about 30 fine-grained task types, for example code:debugging, agent:multi_step_planning, qa_knowledge, math, customer_support or research_report, classified in-flight without requiring retention.
- For the assigned task type, the router looks up which models the OpenRouter community actually spends on over a trailing 7-day window, using the Share of Spend view of the task-spend rankings computed from aggregate anonymized spend statistics.
- OpenRouter describes the ranking as a live signal: when developers migrate a workload to a new model, the router follows within days, with no retraining or manual curation.
- The cost_tier request setting selects one of five cost bands: low, medium, high, xhigh and max, where low favors the cheapest capable models and max favors the most capable regardless of price.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
The figure on the card in front of this brief is not in the material we were given. Neither the Auto Router documentation nor LangSmith's metadata reference states a price, a valuation, or any dollar amount for OpenRouter [cA]. So the $7.5B goes unexplained here, and the mechanism gets the space instead, because the mechanism is what ends up in your traces.
Roughly thirty task types multiplied by the five cost bands the `cost_tier` setting exposes [5] gives about 150 routing buckets [14], each one continuously repriced by people who do not work for you. OpenRouter is direct about the tempo: when developers migrate a workload to a new model, the router follows within days, with no retraining and no manual curation [4].
The measurement consequence is arithmetic rather than opinion. Two evaluation runs seven or more days apart rank candidates from trailing windows that share no data at all [15]. Your repository is unchanged across that gap; the backend is not.
Identification runs backwards too. You learn which model answered from the `model` field in the response [8], while LangSmith will only compute cost when `ls_provider` and `ls_model_name` are supplied together and matched against its pricing database [11]. That pair is an input to your logging whose value now arrives with the reply. The same reference suggests hiding routing and guardrail subspans from the Messages view [12], sensible for transcript hygiene and unhelpful for the one decision you would want to reconstruct later.
The hard edge available is the candidate list: wildcard `allowed_models` patterns, `excluded_models` applied afterwards, and a 404 reading `No models match your request and model restrictions` when nothing survives [13]. Failure is otherwise engineered out, since unavailable classification or rankings fall back to a default model set rather than erroring [7]. Graceful, and one more way the answer moves without telling you.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The Auto Router is invoked by sending openrouter/auto as the model, and works like any other model slug; openrouter/auto-beta is the early-access track.
ReportedView cited source - [2]
A fast, lightweight classifier assigns each prompt one of about 30 fine-grained task types, for example code:debugging, agent:multi_step_planning, qa_knowledge, math, customer_support or research_report, classified in-flight without requiring retention.
ReportedView cited source - [3]
For the assigned task type, the router looks up which models the OpenRouter community actually spends on over a trailing 7-day window, using the Share of Spend view of the task-spend rankings computed from aggregate anonymized spend statistics.
ReportedView cited source - [4]
OpenRouter describes the ranking as a live signal: when developers migrate a workload to a new model, the router follows within days, with no retraining or manual curation.
ReportedView cited source - [5]
The cost_tier request setting selects one of five cost bands: low, medium, high, xhigh and max, where low favors the cheapest capable models and max favors the most capable regardless of price.
ReportedView cited source - [6]
The top surviving models, in market spend-share order, become the primary pick plus fallbacks, after honoring account-level model and provider restrictions, guardrails, ZDR policies, allowed_models restrictions and output-modality requirements.
ReportedView cited source
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- openrouter.ai1h agoAuto Router - Intelligent Model Selection
- docs.langchain.com1h agoMetadata parameters reference

