How we test AI models and how we write our comparisons
Our model pages, comparisons, and guides help people choose between AI models for real work. This page explains what we check, which sources we trust, how we label what is a vendor claim and what is our own output, how we handle our conflict of interest, and how we fix mistakes.
Scope
What we test and write about
NeuroHelper AI gives access to many AI models in one workspace: chat models from OpenAI, Anthropic, Google, xAI, and others in Chat Master; image models such as Midjourney, GPT Image, Nano Banana, Seedream, FLUX, and QWEN Image in Image Master; video models such as Seedance, Wan, Kling, Veo, and MiniMax in Video Master; and voice, music, and sound models in Audio Master. We write three kinds of pages about them:
- Model pages under /models/ explain what one model is built for, what it accepts, and how it works inside our product.
- Comparisons under /comparisons/ put two models or products side by side, or pick the best options for a job.
- Alternatives and guides under /alternatives/ and our guides compare NeuroHelper AI with other platforms and explain workflows.
For every model we look at the same practical questions: what inputs it accepts, what it outputs and at what limits, which jobs it is designed for, where it tends to struggle, and how it fits into a workflow with other models.
Source hierarchy
Which sources we trust, in order
When sources disagree, the higher one wins, and we note the conflict in our source file.
Vendor docs and API references
Model cards, API documentation, release notes, and official pricing pages from the company that makes the model or product.
How the model works in NeuroHelper AI
Which inputs are required, which options are exposed, and which version our app calls. This can differ from the vendor's full API.
Real generations, labeled
Images, videos, and answers we produced in the product, shown with the model entry used. We do not present vendor samples as our own.
Launch posts and vendor blogs
Useful for dates and positioning, but treated as marketing until documentation confirms the details.
Independent reporting
Used for context, never as the only source for a spec or a price.
Rumors and leaked specs
Unconfirmed model names, leaked benchmarks, and social posts are not used as facts.
Labels
How we label every claim
Readers should always know whose claim they are reading.
| Label | What it means | Example wording |
|---|---|---|
| Vendor spec | Documented by the model maker | "Up to 30 seconds (vendor)" |
| Integration detail | How the model behaves in NeuroHelper AI | "In NeuroHelper AI a first frame is required" |
| Our output | A real result we generated | "Real NeuroHelper AI outputs" with the model entry named |
| Official sample | A vendor-made example | "Official xAI samples" |
| Template | A brief you can run yourself, not a result | "Run this prompt with identical inputs" |
| Our view | Editorial judgment, not a measurement | "Best for" and "Watch for" notes |
When a model page uses "(Latest)" in a model name, the exact version is set by the integration and may change. We say so instead of guessing a version number.
Same-brief protocol
How we run a side-by-side test
When we publish outputs from two or more models side by side, we follow the same protocol so the comparison is fair.
Fix the brief
One written prompt, the same reference images or files, and the same target length, ratio, and resolution where both models support it.
Run it in the product
Generate inside NeuroHelper AI with default settings unless a setting is part of the test, and record the model entry and date.
Keep the first results
We show the first usable output or say how many attempts we made. We do not silently pick the best of many runs for one model only.
Judge on stated criteria
Fidelity to the brief, consistency, artifacts, text accuracy, and sound, written down before we look at the results.
When we have not tested: many comparisons start with documentation and briefs before we publish our own outputs. In that case the page says plainly that side-by-side generations are not on the page yet, and the briefs are labeled as templates.
Worked example
How a comparison page comes together
Here is the process behind a typical model-versus-model page, such as a comparison of two video models.
- Inventory check. We confirm which versions are actually available in NeuroHelper AI on the day we write. We compare the versions you can actually run in NeuroHelper AI on the day we write, and we do not imply that a version from another platform is available here.
- Source collection. We read each vendor's documentation and pricing pages and our own model pages, and we save the URLs with the date we checked them.
- Fit mapping. We turn the documented differences into practical questions: which model accepts a last frame, which one makes longer clips, which one needs a starting image, which one generates sound.
- Briefs. We write same-brief tests for the jobs where those differences matter, with criteria for judging them. Until we publish outputs, the briefs are labeled as templates.
- Review. A second pass checks every number against its source, removes anything we cannot trace, and confirms that each link on the page works.
The result is a page that can be wrong only in ways you can check: every claim points to a source with a date, and anything we have not verified is either left out or marked as unverified.
What we do not do
Things you will not find in our comparisons
- No mystery scores. We do not publish 9.2/10 ratings or leaderboards that no one can reproduce. When we recommend a model, we say what job it fits and why.
- No invented results. We never write that we ran a prompt if we did not, and we never describe output quality we have not seen.
- No credit figures in articles. The credit cost of a generation depends on the model, duration, resolution, and inputs, and it can change. The app shows the current estimate before you generate, which is more reliable than a number printed in an article.
- No unverified prices. We quote another company's prices only when we can see them on its official pricing page, and we note the check date. If a price is shown only in an app or behind a sign-in, we say so instead of estimating.
- No guessed versions. If a vendor has released a newer model that is not in our product, we say that rather than implying we offer it.
Conflict of interest
We sell a product, and we say so
NeuroHelper AI is a paid subscription that includes many of the models we write about. That is a conflict of interest when we compare our product with others, and when we decide which models to feature. We handle it in a few concrete ways:
- We describe competitors from their own official pages, with dates and sources you can check. We focus on what you can do in NeuroHelper AI today — and on work already in progress, such as our growing actor and character library — without inventing tests or scores.
- We distinguish what NeuroHelper AI offers from what the vendor's own product offers. We are clear about what you can do with each model in our workspace, based on the integration we offer.
- Calls to action link to our sign-up page and are clearly marked as buttons. The current offer is stated plainly: a $1 trial, then Basic at $19 per month or Pro at $49 per month.
We encourage you to run the same brief in any tool you are considering, including ours, before you pay for a long plan.
Freshness and corrections
How we keep pages current and fix mistakes
Every comparison shows a reviewed date in its methodology block, and every page has a source file that lists the URLs we used and when we checked them. We review pages when a vendor releases a new model or changes pricing, when our integration changes, and on a regular schedule even when nothing has visibly changed.
When we find an error, we correct it on the page and update the reviewed date. If a correction changes a recommendation, we rewrite the affected section rather than leaving a contradiction. When a model is removed from our product or retired by its vendor, we update or redirect its page instead of leaving outdated claims live.
If you notice something wrong or out of date, please tell us through the contact options on our website. Specific links to the source that disagrees with us help us fix it fastest.
Frequently asked questions
Questions about our methodology
Do you test every model you write about?
We document every model from official sources and our integration. We publish our own outputs where we have them, and when we have not run a side-by-side test, the page says so and offers briefs you can run yourself.
Why don't you give models a score?
A single score hides the job. A model that is best for long scenes can be wrong for editing footage. We describe what each model fits and where it struggles instead of publishing numbers we cannot reproduce.
Where do your specs come from?
From the model maker's documentation, API references, and pricing pages, plus how the model works in NeuroHelper AI. We label vendor specs and integration details separately.
Why don't your articles list credit costs?
Credit use depends on the model, duration, resolution, and inputs, and it changes. The app shows the current estimate before you generate, which is more reliable than a number in an article.
Are your comparisons biased toward NeuroHelper AI?
We sell a subscription, and we say so. We cite competitors from their own pages, keep vendor claims and our product separate, and write about the versions and tools you can actually use in NeuroHelper AI.
How often are pages updated?
When a vendor releases a model or changes pricing, when our integration changes, and on a regular review schedule. Each comparison shows its reviewed date, and each page has a dated source list.
What happens when you find a mistake?
We correct the page, update the reviewed date, and rewrite any recommendation the error affected. We do not leave contradictory text in place.
Can I run the same tests myself?
Yes. Our comparisons include copyable briefs. Run them with identical inputs in each tool you are considering and judge the results against the criteria listed with each brief.
How this comparison was built
Where this methodology applies
This page describes how we build model pages, comparisons, alternatives, and guides on neurohelper.ai.
Model pages
One model: inputs, outputs, limits, jobs, and real or official examples, clearly labeled.
Comparisons
Two models or products, or the best options for a job, with sources and templates for your own tests.
Guides and alternatives
Workflows and platform comparisons, including where other tools fit better.
See for yourself
Run your own same-brief test
Use leading chat, image, video, and audio models in one workspace and compare them on your own work.
