CodeSOTA / Task Router

Your task stays the same.
The model can change.

Choose the work, compare candidates, and inspect the route through a common API. We are building toward one integration across documents, speech, code, text, and vision.

Model selection is public. The code and text execution API requires a signed-in account and a configured provider. Other task families currently offer model selection.

One selection API, many tasks.

This panel queries the registry. No inference or credit charges occur here.

Benchmark tradeoffs

Find the Pareto frontier.

Every dot is a recorded result. Highlighted models have no competitor in this cohort that is at least as good on both axes and strictly better on one.

Loading benchmark observations…

01 / Define the work

Documents

Input
Scans & PDFs
Target output
Text, tables & fields

Registry selection only · execution planned

Target outputs describe the task. This preview returns model recommendations. It does not process your files or run inference.

Explore documents evidence →

02 / Inspect the shortlist

03 / Use the same selection API

curl 'https://www.codesota.com/api/pareto-router?task=document-ocr&objective=balanced&limit=3'

Direction / Task contracts

A stable result format
for each kind of work.

A document parser, a speech model, and a coding model need different inputs and outputs. The roadmap keeps those contracts explicit while sharing model discovery, routing, accounts, and usage reporting.

See the current API contract →
  1. Selection, available now.GET /api/pareto-router returns candidates, alternatives, recorded evidence, and a cost basis for registry tasks.
  2. Code and text execution, implemented.POST a prompt or structured input from a signed-in session. The configured provider route returns generated text and account metadata. Provider availability needs workload validation.
  3. More execution families, planned.Document parsing first, then speech and vision adapters with versioned recipes, validated task outputs, asynchronous jobs, and fallback policies.

How to read the result

Inspect the evidence
behind the shortlist.

The registry combines reported benchmark results. Selection scores aggregate those records; costs may be inferred, and results can come from different evaluation protocols. Use the original source and representative inputs to assess fit.

Evidence methodology →
  1. Quality first: best.Sort by the registry quality score.
  2. Balance: balanced.Combine quality and inferred affordability.
  3. Cost first: cheap.Prioritize the affordability heuristic. This is not a verified price ceiling or measured quality floor.