Tests, comparisons & original research

How We Know

Independent tests where we show the setup, keep the mistakes and explain exactly how we reached the verdict.

Featured test

Best AI Image Generators in 2026: Tested and Ranked

We give every major AI image generator the same three prompts, keep the first image and rank the results on photorealism, on-screen text and product advertising. The ranking is updated as new models arrive.

Read the full test

All Tests & Research

Eighteen AI coding harness logos arranged as floating glass medallions on a warm ivory surface, with DeepSeek Harness shown most prominently at the centre.

AI & Tech

Best AI Coding Harnesses 2026: TGK Top 10 and Tier List

DeepSeek Harness won our TGK discussion, while Hermes divided the room. Here are our Top 10, the full 18-tool tier list and what the benchmarks show.

Ten major AI model and company logos floating as ivory medallions in a warm cream, high-tech editorial composition.

AI & Tech

Best AI Models in 2026: TGK's Top 10 and Full Tier List

Opus 5.5 leads our Top 10, Astra is close behind and GPT-6 Sol is our everyday pick. Here's which AI model to use for what, which ones burn the most tokens, and our full 39-model tier list.

A dark green rounded-square app icon with a white “u.” centered on a soft cream background.

Things to Use

Can a Non-Developer Build an App With Codex?

We took a clickable prototype into Codex and built a live web app. Then we tried using it, found bugs the tests had missed and showed it to three people.

A clean grid of small identical blank rectangular cards, evenly spaced and evenly lit on a seamless warm ivory surface

AI & Tech

How to Compare AI Models: The TGK Testing Method

This is the method TGK uses to compare AI models without hiding rerolls, failed attempts, different settings or the limits of each test.

Three action figures struggle in a miniature hotel lobby while their hands blur together and a red gift remains sharp beside the elevator.

AI & Tech

AI Fight Scene Test: How to Write a Better Prompt

We rewrote one difficult AI fight prompt and tested it in Seedance, MiniMax and FLUX 3 to see what better structure fixed and what it could not.

Five play button icons arranged from small matte plastic to larger brushed metal and polished glass versions on a clean white studio surface.

AI & Tech

How Much Does a Usable 10-Second AI Video Cost?

One AI video can cost $1, but one usable clip usually costs more. Here is what we paid, what failed and how much to budget for rerolls.

AI-generated hotel fight comparison with a suited man holding a red gift as attackers and elevator doors behave differently across five video models.

AI & Tech

How 5 Popular AI Video Generators Handled Our Tests

Five popular AI video generators received the same conversation, hotel-fight and drink-ad prompts. We kept every first result and every strange mistake.

Grok 4.6 vs Claude Opus 5 vs GPT-5.6 Sol across five tests of cost, speed, writing, coding and visual judgment.

AI & Tech

Can Grok 4.6 Beat Claude Opus 5 and GPT-5.6 Sol? We Tested It

We put Grok 4.6 against Claude Opus 5 and GPT-5.6 Sol in five real tasks. It did not win the set, but it proved it can call the best models its peers while charging much less.

Four AI-generated video frames comparing Seedance, Kling, Veo and Wan on portrait, walking and perfume-ad tests.

AI & Tech

Best AI Image-to-Video Generators in 2026: 4 Popular Models Tested

Seedance made the best AI image-to-video clips in my three tests. Kling came surprisingly close for around one-third of the price.

AI-generated portrait, tennis and jump-rope test scenes arranged as video storyboard frames for a Seedance 2.5 review.

AI & Tech

Seedance 2.5 Review: Brilliant for 5 Seconds, Sloppy at 30

Five Seedance 2.5 tests cost us only $19.62. Some clips looked almost real. Then our 30-second tennis match lost the ball and the plot.

Split-screen comparison showing a Mage-Flow Fast mode café image beside the Quality mode error screen from Microsoft’s Hugging Face demo.

AI & Tech

Microsoft Mage-Flow Test: Fast Mode Was Fast. Quality Mode Never Worked

I tested Microsoft’s new Mage-Flow image model through its public Hugging Face demo. Fast mode produced an ugly, obviously AI image; Quality mode produced nothing at all and still burned through my credits.

AI & Tech

Nano Banana Pro Prompting Guide: How to Get Better Results Without Overcomplicating Your Prompts

Long Nano Banana prompts can work perfectly, but most images do not need one. We tested both approaches and looked at which instructions made a difference.

AI & Tech

How to Create a Consistent AI Character Without LoRA Training

TGK tested a four-image Nano Banana Pro process for keeping the same AI face and body across completely different scenes without training a LoRA.

Comparison of first results from seven AI image generators using the same Mediterranean café prompt, with Nano Banana 2 marked as the winner after 36 editing tests.

AI & Tech

We Tested 7 AI Image Editors Across 36 First-Try Edits

I gave seven AI image generators the same task, then tested six of them with the same six edits. Every image shown is the first result I got.

A traveller works on a laptop in front of a world map surrounded by pinned photos from France, Morocco, Colombia, Cambodia, Russia and Iraq, each showing a different VPN service with varying

Things to Use

I’ve Used VPNs All Over the World. Here’s What Still Works

NordVPN, ExpressVPN and Surfshark are decent when you want something simple. In more difficult countries, the protocol and server setup matter far more than the famous logo.

Data & citations

Research Archive

Download the prompts, model settings, recorded costs and failure notes behind our largest AI comparisons.

AI Image Generators 2026

Thirteen models tested on three shared briefs in August and October 2026, with published places, prompts, recorded costs, output links and documented exceptions.

13 models · 3 shared briefs · CC BY 4.0, artwork excluded

Browse dataset Download ZIP Read full test

AI Video Generators 2026

Five models, three shared tasks, 15 first-result observations and a separate action-prompt retest.

5 models · 3 shared tasks · CC BY 4.0

Browse dataset Read full test

Method and limitations

These are small editorial tests. The original video prompts were reconstructed from our testing record rather than exported from the generation interfaces. Models ran at different available resolutions, and we retained one completed result per model and task, so the archive documents this comparison rather than repeat-run reliability.