Auto Model Routing

Set the model to auto and BLACKBOX AI picks a closed, big open-weight or small open-weight model for every request, based on what the request needs.

Coding agents make one model call per step, and most steps (listing files, rerunning a test, reading output) do not need a closed model. With model: "auto", BLACKBOX AI reads each request, decides which tier of model the step needs, and serves it with that model. Hard steps go to a closed model; moderate and routine steps go to lower-cost open-weight models.

Routing is decided per request. Agents resend the full conversation on every turn, so whichever model handles a step sees everything earlier steps did.

Quick start

Use the model ID blackboxai/auto (or simply auto) with any supported endpoint. Everything else in the request stays the same.

curl https://enterprise.blackbox.ai/chat/completions \
  -H "Authorization: Bearer $BLACKBOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "blackboxai/auto",
    "messages": [
      {"role": "system", "content": "You are a coding agent."},
      {"role": "user", "content": "The date parser test fails. Find and fix the bug."}
    ]
  }'

import os
import openai

client = openai.OpenAI(
    api_key=os.environ.get("BLACKBOX_API_KEY"),
    base_url="https://enterprise.blackbox.ai",
)

response = client.chat.completions.with_raw_response.create(
    model="blackboxai/auto",
    messages=[
        {"role": "system", "content": "You are a coding agent."},
        {"role": "user", "content": "The date parser test fails. Find and fix the bug."},
    ],
)

print(response.headers.get("x-blackbox-routed-model"))  # the model that served this step
print(response.parse().choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.BLACKBOX_API_KEY,
  baseURL: "https://enterprise.blackbox.ai",
});

const { data, response } = await client.chat.completions
  .create({
    model: "blackboxai/auto",
    messages: [
      { role: "system", content: "You are a coding agent." },
      { role: "user", content: "The date parser test fails. Find and fix the bug." },
    ],
  })
  .withResponse();

console.log(response.headers.get("x-blackbox-routed-model")); // the model that served this step
console.log(data.choices[0].message.content);

To route a coding agent, set its model to blackboxai/auto and its base URL to https://enterprise.blackbox.ai. No other change is needed.

Supported endpoints

Once the model is chosen, the request continues through the normal pipeline as a request to that model, including tool calls and safety checks.

Model tiers

Each request is served by one of three tiers:

Which model served the request

Every routed response carries three headers:

How the decision is made

For each request, BLACKBOX AI builds a short summary of the agent's state: the task, the most recent messages and tool output, and counters such as the turn number and how many commands have failed so far. Microsoft Decision-1 answers a fixed set of yes/no questions about that state in a single call (for example, "Did the last command fail?" or "Will the next action edit source code?"), and a small learned model turns the answers into a tier. A routing decision typically adds under a second.

On held-out evaluations across coding agents including Claude Code, Codex, Gemini CLI, OpenHands and SWE-agent, the router picks the same tier as a GPT-6 Astra judge 70 to 80% of the time. See Blackbox AI Router, Powered by Microsoft Decision-1, Outperforms Claude Opus 5.5 on Terminal-Bench for the full evaluation.

Reliability

Routing never fails a request. If the router cannot decide in time or is unavailable, the request is served by the big open-weight tier (Kimi K3), and x-blackbox-route-source is default.

Best practices

• Use it for multi-step agents. The savings come from routine steps, which make up most of a coding agent's calls.
• Pin a model when you need one. If a request must be served by a specific model, send that model ID instead of auto.
• Log the routing headers. Recording x-blackbox-routed-model per request shows how your traffic is split across tiers.