Evaligo is an AI workflow builder with built-in A/B testing. You build a flow of prompts, agents and tools, compare models, prompts or whole flows on the same inputs with blind AI judges, and deploy the winning flow as an API.

Can Evaligo pick a model for each input?

Yes. A finished experiment can train a router (the KNN method from UIUC's LLMRouter library) that picks a model per input. In a published listing-localization test it scored 0.909 on 28 unseen listings versus 0.911 for Claude Sonnet 5, at $0.00033 versus $0.00885 per listing.

Is Evaligo free to use?

Yes, Evaligo offers a free tier that lets you build and deploy AI workflows without a credit card. The free tier includes generous limits for personal projects and testing.

Do I need to know how to code to use Evaligo?

No coding is required. Evaligo uses a visual drag-and-drop interface where you connect nodes to build workflows. However, developers can also use custom code nodes for advanced use cases.

What can I build with Evaligo?

You can build AI agents, chatbots, content generators, data pipelines, automation workflows, and AI-powered apps. Common use cases include marketing automation, customer support bots, content creation, and data processing.

Visual AI workflow builder with A/B testing for models and prompts

Build a flow of prompts, agents and tools. Run it on several AI models, score every output with blind AI judges, and compare quality, cost and speed. Then deploy any flow as an API.

Drag-and-drop flow editorAny model, incl. open sourceBlind AI judge panelsQuality, cost and speed per outputModel routing per inputDeploy any flow as an API

Run your first A/B test free See a real benchmark

Evaligo A/B test scoreboard: six AI models compared on the same 30 inputs, scored by three blind judges, with cost and speed per output

A real Evaligo experiment: six models, the same 30 inputs, three blind judges. Read the full write-up.

What testing found on our own workflows

Three experiments run on Evaligo. Each ran every model on the same inputs and recorded the score and the cost per output.

Language detection · 129 texts

Six models, the same score, 32× the price

Every model scored 0.99 or better on the same 129 texts. The most expensive one cost 32 times the cheapest and did not score higher.

Model	Score	Cost
gpt-5.4-mini	1.000	$0.00015
claude-sonnet-4-6	1.000	$0.00063
claude-haiku-4-5	0.992	$0.00023
gemini-3.5-flash	0.992	$0.00485

Scored against a gold answer set. Cost is per text.

Listing localization · 63 listings, 7 languages

0.006 apart, 20× apart in cost

Claude Sonnet 5 scored highest, but gpt-5.6-luna came within 0.006 of it for a twentieth of the price per listing.

Model	Score	Cost
claude-sonnet-5	0.914	$0.00906
gpt-5.6-luna	0.908	$0.00045
gemini-3.7-flash	0.897	$0.00328
deepseek-v4-flash	0.893	$0.00021

Blind three-judge panel. Cost is per listing.

Read the case study

Writing level · 87 texts

The cheaper model graded better

Flash scored higher than the Pro model of the same family, and cost less per text. Price did not predict quality.

Model	Score	Cost
gemini-3.5-flash	0.977	$0.00648
gemini-3.1-pro	0.948	$0.00774
gemini-3-flash	0.937	$0.00175
gemini-2.5-pro	0.845	$0.01151

Graded against a reference answer set. Cost is per text.

Compare models, prompts and flows on the same inputs

Drag-and-drop flow builder

Describe the flow in one sentence and Evaligo builds it: prompts, web scraping, data extraction, tools and agents as connected nodes. Drag nodes on the canvas to edit it, or change it by instruction.

View Docs Build a flow free

Drag-and-drop flow builder - Evaligo A/B testing

Any model as a variant, including open source

Pick the models you want to compare: OpenAI, Google, Anthropic, and open-source models such as DeepSeek, MiniMax, Kimi and Qwen through one OpenRouter key. Each variant runs the same flow on the same inputs; only the model, prompt or reasoning setting changes.

View Docs Compare models free

Test on your own data, or generate it

Run every variant over a dataset of real inputs. If you have none yet, Evaligo generates a realistic test set from a one-line description, so a 30-sample comparison does not wait for you to collect data.

View Docs Create a test set

Test on your own data, or generate it - Evaligo A/B testing

Judge every output blind

Scoring rubrics, written once

Write the criteria once: what a good output must do, and what costs points. AI judges score every output against the rubric without knowing which variant produced it, and record a one-line reason per score.

View Docs Write your first rubric

Scoring rubrics, written once - Evaligo A/B testing

Judge panels from several vendors

Use one judge model or a panel of three from different vendors, averaged. Per-judge scores stay visible, so you can see where the judges agree and where they do not.

View Docs Set up a judge panel

Repeat runs

Model output and judging both vary between runs. Re-run an experiment with one click and compare runs side by side before you decide.

View Docs Run an experiment

Read the scoreboard and deploy

Quality, cost and speed per output

Every experiment ends in one table: score, flawless outputs, cost per output and latency for each variant, with a score distribution per sample. Export it as a PDF report.

View Docs

Promote the winning variant

Make the winning model, prompt or flow the live version with one click. The flow keeps its inputs, outputs and API endpoint.

View Docs

Train a router on the results

When an experiment finishes, one click trains a router on its scoreboard. It learns which model handles which kind of input, shows quality against cost with 95% error bars so you can choose a working point, and then runs as a model in your next experiment or in a deployed flow.

Read the case study

Deploy as a production API

Any flow deploys as a REST API with authentication, documentation and code examples. Batch processing and monitoring are built in.

View Docs

Benchmarks we published, run on Evaligo

Every number below comes from an experiment you can reproduce: the same flow, the same inputs, blind judges, cost and speed per output.

Open source vs closed on writing

Six models rewrite 30 dating bios; three blind judges. gpt-5.6-luna 0.883, gemini-3.7-flash 0.879, DeepSeek V4 Pro 0.849.

Read the scoreboard

Model versions, not just vendors

Salary negotiation coach on 30 offers. Gemini Flash 3.5 scored 0.68; the 3.7 release scored 0.75 at 3.5x lower cost, with no prompt change.

Read the scoreboard

Agent architectures head to head

One agent, an assembly line, or agent plus editor: the same task built three ways, eight model mixes, a 14-criteria judge panel.

Read the scoreboard

Model routing

Router: 0.909 vs Sonnet’s 0.911, at 4% of the cost

In the listing localization test, a router trained on the four-model scoreboard picked a model for each of 28 listings it had not seen. It scored 0.909, Claude Sonnet 5 scored 0.911, at $0.00033 versus $0.00885 per listing.

Any finished experiment can train a router. You pick the quality and cost trade-off with 95% error bars, and the router runs in your next experiment or flow like any other model.

Read the case study Train a router free

Listing localization, 28 held-out listings, judge score and cost per listing

Claude Sonnet 5 (every input)0.911

$0.00885 per listing

Trained router (picks a model per input)0.909

$0.00033 per listing

Bar length is cost. Scores from a blind three-judge panel. Method: KNN router from UIUC’s open-source LLMRouter library.