Sources and evidence

Product documentation explains the workflow. User reviews describe experiences. Independent benchmarks evaluate models and tasks. Read each source within its scope.

How to read the evidence

Product pages describe documented capabilities and intended workflows; they are not hands-on test reports.
G2 and Capterra scores are shown separately, with the review count, original source and observation date. They describe the product listed on each platform.
A benchmark result applies to the tested model version, task, and evaluation method. Using that model does not give a tool the same result.
Compare pricing, limits, and supported features on the official site before choosing a plan.

How to read the evidence

Benchmarks

Artificial Analysis

Artificial Analysis · video evaluations

Artificial Analysis · Video generation and editing models. What it measures: Viewing preference. Text, image and video inputs use separate evaluations. Audio and silent outputs are also evaluated separately.

Scope: Video generation and editing models
What it measures: Viewing preference

VBench team

VBench Leaderboard

VBench team · Video generation models. What it measures: Quality and consistency dimensions. Examines video quality across 16 dimensions, including subject consistency, motion and temporal behavior.

Scope: Video generation models
What it measures: Quality and consistency dimensions

Artificial Analysis

Artificial Analysis · speech evaluations

Artificial Analysis · Text-to-speech models. What it measures: Listening preference and pronunciation. Listening, pronunciation and API speed are separate measures. Language and voice selection affect comparisons.

Scope: Text-to-speech models
What it measures: Listening preference and pronunciation

TTS-AGI & contributors

Speech synthesis evaluation sources

TTS-AGI & contributors · English text-to-speech prompts. What it measures: Anonymous listening comparisons. Listeners compare two anonymous voices reading the same text. Current prompts are English only.

Scope: English text-to-speech prompts
What it measures: Anonymous listening comparisons

BigCode project

BigCodeBench

BigCode project · Function-level code generation. What it measures: Programming task pass rate. Complete and Instruct use different prompts; Full and Hard use different task sets. Match them before comparing results.

Scope: Function-level code generation
What it measures: Programming task pass rate

Artificial Analysis

Artificial Analysis · image evaluations

Artificial Analysis · Image generation and editing models. What it measures: Viewing preference. Text-to-image and image editing are separate evaluations. Editing tests changes to an existing image.

Scope: Image generation and editing models
What it measures: Viewing preference

Hugging Face & contributors

Open ASR Leaderboard

Hugging Face & contributors · Speech recognition systems. What it measures: Word error rate and RTFx. Separate tracks cover short and long English recordings and short multilingual recordings. RTFx measures transcription speed relative to recording duration.

Scope: Speech recognition systems
What it measures: Word error rate and RTFx

Datacurve

DeepSWE

Datacurve · Coding models in the evaluator’s agent. What it measures: Repository task completion. Version 1.1 grades committed code changes in a separate environment. Results depend on the tested agent and reasoning configuration.

Scope: Coding models in the evaluator’s agent
What it measures: Repository task completion

Product artwork

Product names, logos and promotional images identify the respective tools and belong to their owners. These links show the official pages from which the identities were sourced.

Midjourney is a trademark of Midjourney, Inc. We are not endorsed by or affiliated with Midjourney, Inc.