---
title: "AI Generation Research: Cost Evidence and Model Benchmarks | Made"
description: "Dated generation-cost evidence with prompts, every attempt and downloadable data, plus five preregistered AI model studies awaiting blinded human ratings."
url: https://www.made.app/research
---

Research

# AI generation research you can audit

Read the measured costs and outputs from a completed comparison, or inspect the frozen protocols for our five planned model studies. Each report states its evidence, scope and limits. The model studies remain preregistered until their required blinded ratings are complete.

## Completed observational reports

Observed 2026-09-23 · Unblinded cost evidence

### [Made and Higgsfield: a matched-output cost comparison](https://www.made.app/research/made-higgsfield-matched-output-cost-2026-09-23)

Eighteen completed generations with matched named models, prompts, source files and measured native specifications. Made used $5.88 of historical wallet balance; Higgsfield used 123 credits, or $5.945 under the disclosed full-plan-utilization allocation.

Read the report and inspect every attempt

## Requirements for the five planned model studies

The protocols below were frozen on 2026-08-13. These rating requirements apply to those model studies.

- **Preregistered.** Prompt sets are frozen at version 2026-q3-v1 and published as JSON before testing starts, so prompts can't be tuned toward a preferred winner.
- **Blinded.** Two raters score every output with model labels and ordering randomized per prompt; both scores and their disagreement are published.
- **No cherry-picking.** Provider errors, refusals, and corrupt outputs count as failures — reruns beyond the preregistered count are prohibited.
- **No sponsors.** No provider pays for compute, reviews scores before publication, or influences the ranking. Testing runs through Made's ordinary production routing.

## Preregistered model studies

### [Text-to-Video Model Comparison](https://www.made.app/research/text-to-video-model-comparison-2026-q3)

Preregistered

WAN 2.2 Turbo, Veo 3.1, Kling 3 Turbo — 20 frozen prompts, 3 runs each, results publish with raw data when complete.

### [Image-to-Video Motion and Identity Test](https://www.made.app/research/image-to-video-motion-identity-2026-q3)

Preregistered

WAN 2.2 Turbo, Kling 2.5 Turbo, Veo 3.1 — 20 frozen prompts, 3 runs each, results publish with raw data when complete.

### [Face and Character Consistency Test](https://www.made.app/research/face-character-consistency-2026-q3)

Preregistered

Nano Banana, Nano Banana 2, Nano Banana Pro — 20 frozen prompts, 3 runs each, results publish with raw data when complete.

### [Image-Model Prompt Adherence Test](https://www.made.app/research/image-model-prompt-adherence-2026-q3)

Preregistered

Flux 1.1 Pro, Nano Banana, Nano Banana 2, Nano Banana Pro — 20 frozen prompts, 3 runs each, results publish with raw data when complete.

### [Cost and Time per Usable Output](https://www.made.app/research/cost-time-per-usable-output-2026-q3)

Preregistered

WAN 2.2 Turbo, Veo 3.1, Flux 1.1 Pro, Nano Banana, Nano Banana 2, Nano Banana Pro — 40 frozen prompts, 3 runs each, results publish with raw data when complete.

The models under test run in Made — see the [price sheet](https://www.made.app/price-sheet) for exact per-generation costs, or [spec-level comparisons](https://www.made.app/compare) while the benchmark results are pending.

---

Markdown version of https://www.made.app/research. Every public Made page has one at its URL plus .md (https://www.made.app/research.md). Site summary for AI assistants: https://www.made.app/llms.txt