Auto Exacto

Not yet available.

Auto Exacto is OpenDunes' accuracy-first routing mode. Instead of you picking a model, you describe a task category and OpenDunes selects the highest-accuracy model available within your optional budget ceiling — factoring in live benchmark scores, provider reliability, and per-token DA cost.

How it works#

Set model to the special value "opendunes/auto-exacto". OpenDunes evaluates the candidates from the full catalog against your constraints and routes to the one scored highest for accuracy on that task type.

{
  "model": "opendunes/auto-exacto",
  "messages": [{ "role": "user", "content": "Prove that sqrt(2) is irrational." }]
}

The model field in the response tells you which model actually ran. The selection is deterministic for a given catalog snapshot — if catalog scores change, routing may change.

Accuracy scoring#

OpenDunes maintains an internal accuracy index per model, updated as providers report performance data. The index weighs:

• Standardized benchmark results (reasoning, coding, instruction-following)
• Observed output quality scores from anonymized request sampling
• Provider reliability — an accurate model that times out half the time scores lower

You cannot inspect the raw scores directly, but you can see the current top-ranked model for a task type by calling GET /api/models?sort=accuracy (not yet available; ships alongside this feature).

Budget ceiling#

To keep Auto Exacto within a DA spend per request, pass max_cost_da:

{
  "model": "opendunes/auto-exacto",
  "max_cost_da": 0.5,
  "messages": [{ "role": "user", "content": "Write a cover letter in French." }]
}

OpenDunes filters the candidate set to models whose estimated per-request cost (based on your message length and max_tokens) is at or below the ceiling, then picks the most accurate among those.

If no model fits within the ceiling, the request fails with 402 insufficient_credits and a message indicating the cheapest available model's estimated cost.

Task hints#

For better selection, pass a task hint in the routing object:

{
  "model": "opendunes/auto-exacto",
  "routing": { "task": "reasoning" },
  "messages": [{ "role": "user", "content": "Solve: 3x + 7 = 22" }]
}

Pricing (DA)#

Auto Exacto routes to whichever model wins the selection — you pay that model's normal DA rate. The cost is not fixed. For predictable billing, combine max_cost_da with your budget, or use a specific model slug instead.

Using with the OpenAI SDK#

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://opendunes.com/api/v1",
    api_key=os.environ["OPENDUNES_API_KEY"],
)

resp = client.chat.completions.create(
    model="opendunes/auto-exacto",
    messages=[{"role": "user", "content": "Explain Dijkstra's algorithm step by step."}],
    extra_body={
        "routing": { "task": "reasoning" },
        "max_cost_da": 1.0,
    },
)
print(resp.choices[0].message.content)
print("Routed to:", resp.model)

Note. Auto Exacto is distinct from the Exacto variant (:exacto suffix on a specific model slug). Auto Exacto picks the model; the Exacto variant uses accuracy-optimized settings on a model you already chose.