DigitalOcean's Inference Router now picks the best model per request

One model name, many models behind it

Developers set the model to inference-router, send a prompt and, if they want, add a preference. The router then forwards each request to the model it judges best for that job, validated by public benchmarks, according to DigitalOcean.

Behind the router sits the full DigitalOcean Model Catalog. That includes frontier models from Anthropic and OpenAI as well as open source models such as Kimi, GLM and DeepSeek. DigitalOcean says every new model gets the same test, and those that measure up are added to the router.

The company frames the problem as one of constant churn: a new model claims to be best in class every few weeks, and teams spend a week on evals, side-by-side tests and debate before deciding whether to switch. The router is meant to take over that decision. DigitalOcean also argues that routing each request to a fitting model means customers stop paying frontier prices for work a leaner model can handle just as well.

In its internal testing across software engineering, reasoning and agentic work, DigitalOcean says Inference Router had higher accuracy than Claude Opus 5.5 and GPT-6 Astra, at about 22% lower cost. The company notes that its scores come from internal evaluations as of October 12th, 2026, and that results vary with configuration and may differ from public leaderboards.

How the routing works

DigitalOcean says it measures models across a wide variety of tasks, scoring quality along with cost and latency. That evidence decides which model wins each task. When a request arrives, a routing model reads it, matches it to a task and weighs the customer's preference to pick a model.

The router is described as cache-aware. It tries to stay with one model to reuse its cache, unless switching pays for itself. Because the evaluation work happens before a request arrives, customers do not need to manage configuration by hand, and a newly released model that does better on a given task is picked up automatically.

The router is initially trained on three task categories:

• Software engineering
• Agentic workflows
• General intelligence and reasoning

The Getting Started page in the DigitalOcean cloud console shows the benchmarks behind each use case, how the router scores against the best single model and which models it chooses between. According to the company, the page updates whenever a model is added or an evaluation is rerun.

Preferences and custom evaluation sets

Most requests should need no input, DigitalOcean says, as the service is managed end to end by the company. Users who want to steer the outcome can use the budget preference to favor lower cost models, or quality for cases where success is critical. The default, routing_preference: balanced, is expected to work for most customers.

DigitalOcean acknowledges that strong results on public benchmarks and internal evals do not show how the router performs on a customer's own workload. For that reason it plans to let users bring their own rubric. Customers would supply an evaluation set, and their definition of success would decide which model wins. The company says it will rerun those evaluations when a new model ships. This feature is not yet available, and early access can be requested from the Getting Started page in the DigitalOcean Console.

Availability and how to start

Inference Router is available today in Public Preview. DigitalOcean says there is no additional cost during the preview, and users are billed only for the models that serve each request.

To get started, open Inference Router in the DigitalOcean Control Panel and generate a model access key if one is not yet configured. API calls then go to https://inference.do-ai.run/v1 with inference-router as the model. The example in the announcement is a chat completions request with the preference set to balanced.

For teams that now run their own evaluations and routing rules every time the model landscape shifts, the pitch is a single endpoint that handles the choice. Whether the claimed accuracy and cost gains hold for a given workload remains to be tested, which is why the planned custom evaluation option matters.