World Model Optimizer
GitHub
World Model Optimizer (WMO) is a new open-source toolkit that turns AI agent logs into smaller, cheaper models. It routes requests between big and small models, matching top-tier quality for less money.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There's a new open-source tool called World Model Optimizer, or WMO, and its pitch is refreshingly practical: stop paying frontier prices for every single request your AI agent makes. Instead, WMO collects the traces your agent already generates, uses them to train smaller open-source models through the Tinker API, and then routes traffic intelligently between the big frontier model and the smaller ones depending on what a given task actually needs.
The routing piece is where the numbers get interesting. On RouterBench, WMO's serving layer held onto frontier-level quality while cutting costs by 27%. That's not a marginal tweak — it suggests a meaningful chunk of agent traffic doesn't need the most expensive model to get a good answer, and WMO is built to figure out which requests those are automatically, using a policy fit on your own held-out traces rather than someone else's benchmark.
The setup process is refreshingly hands-on for anyone who's dealt with router configs before. You register your model providers, and WMO goes and pulls their full catalogs — OpenRouter alone contributes 338 published models to the pool it can choose from. From there you build a model from your OpenTelemetry traces, sweep every registered model against held-out tasks, and fit a routing policy using something like a k-nearest-neighbors approach, deterministically splitting your traces 30/70 between reporting and fitting so the evaluation doesn't cheat by peeking at its own training data.
Beyond routing, WMO ships with actual world models — simulated environments you can step through programmatically or over HTTP, useful for testing an agent's behavior without touching production systems or racking up API bills on real calls. There's also a hosted platform version that runs agents inside E2B sandboxes, keeps model credentials off the sandbox entirely, and lets an optimizer rewrite prompts, tools, and even runtime code, with changes only becoming the new default once they clear evaluation gates. Long-running jobs can detach and reattach later, which matters if you're running something that takes hours rather than seconds.
One detail worth flagging for anyone wary of silent data collection: WMO has anonymous usage telemetry on by default, but the project is explicit that it captures only metadata — no prompts, no traces, no file paths, no credentials. It's opt-outable with a single config command or an environment variable, which is the kind of transparency that's still rarer than it should be in dev tooling.
My take — AI-written commentary, not fact-checked reporting
The 27% cost cut on RouterBench is the headline, but the more interesting bet here is that agent traces are a renewable resource — rerun the pipeline as new traces come in and the model you own keeps getting better, which is a very different economic model than just renting a frontier API forever. Telemetry defaulting to on will annoy some people regardless of how metadata-only it claims to be, and that's a fair fight to have even when the intentions are good. Open tooling that lets teams train their own smaller models off their own usage data, instead of staying permanently dependent on one frontier provider, is exactly the direction this space should be moving.}}
Read more about this at: GitHub