Zone 6 · My own tests

TestLAB: my own tests, benchmarks and rankings

This is where I publish my own measurements: the same tasks, the same method, raw data. Alongside, an external benchmark for reference: the Artificial Analysis Intelligence Index.

6planned tests

3×repeats of every test

0results without a method

Suggest a test

Writing and workshops

From my knowledge base

GitHubOct 20Inspect: overview, install and notes

Inspect is a framework for evaluating models and agents from the UK AI Security Institute. I cover its building blocks, ready evaluations and when to use it.

GitHubOct 20llama.cpp: overview, install and notes

llama.cpp is an inference engine in C/C++ for running models locally. I cover installation, the README commands and when to pick it over simpler tools.

GitHubOct 20Ollama: overview, install and notes

Ollama is a tool for running open models on your own computer with a model library and an API. I show the installation, first commands and context setting.

Shop499 złAI Marketing Agent 1:1

Wybierz powtarzalne zadanie i sprawdź, jak uprościć je z AI. Indywidualny warsztat online, 90 minut.

AI News from my base

News on this topic

AI NewsOct 25A new AI model with a huge memory is ready to try today

Want to test a fresh AI model without opening yet another account with its maker? I explain how to start it, what the maker promises and what is still unknown.

AI NewsOct 24Check in one command how much you have spent on Vercel

A hosting bill can surprise you, especially when agents work for you. I show how to check spending from the terminal and set a budget that can pause sites.

AI NewsOct 23A small AI model can cost 77% less if you trim its vocabulary

Want to run your own AI model more cheaply? Dharma-AI says trimming its vocabulary cut cost by over three quarters. I explain which parts to trust.

AI NewsOct 20Build an iPhone AI app once, swap the model without rewriting

Building an AI app for iPhone or Mac? One new tool lets you swap the model for another without rewriting your code. I show what it means in practice.

Topics to dig into

Hubs in this zone

In this zone

My own measurements

TestLAB: method and test plan

I do not publish results without a method. First a plan and rules, then measurements, the ranking last. Every test gets a page with prompts, model versions, date and raw data.

01The same tasksEvery model and tool gets an identical brief under the same conditions.
02Three repeatsThe result is the median of three runs, and the spread stays visible.
03Cost and timeI count tokens, minutes and corrections, not just an impression of quality.
04Raw dataPrompts, model versions, the test date and the results are published together with the ranking.

CreatePlan

The same brief, five models: a site from a five-sentence brief

How the first version of a site differs for the same project description.

Method: 3 runs, brief from the blog
We measure: time, cost, number of fixes

CreatePlan

Claude Code, Codex CLI and Gemini CLI: refactoring a small repository

Three coding agents on the same task and the same repository.

Method: same repo, same prompt
We measure: tests passed, diff, cost

CreatePlan

Local models on my hardware: tokens per second

What open-weight models running locally really deliver.

Method: fixed hardware, 3 models
We measure: tokens/s, memory, quality

AutomatePlan

An agent with human control: approvals per hour of work

Whether the level of autonomy from the “Human in the loop” post holds up in practice.

Method: the warehouse-agent pilot
We measure: approvals, errors, time

ScalePlan

The hook in the first three seconds: ten versions

A test of one format with different openings on one account.

Method: one format, 10 versions
We measure: reach, retention, saves

ScalePlan

Content from one customer question: how many posts from two hours

Putting the weekly content rhythm to the test.

Method: 4 weeks, the same rhythm
We measure: number of posts, time

TestLAB ranking
It will appear after the first three tests are finished. Until then the reference point is the external metric above, marked as Artificial Analysis data.

Suggest a test Notify me about results

Frontier model ranking

Artificial Analysis Intelligence Index

The best on the left, the weaker on the right. Click a bar for details. Change the metric or show only open-weight models.

Data: Artificial Analysis · Oct 19, 2026

ClosedOpen weightsSelected

Claude Opus 5.5Anthropic · level Max · closed model

57.6 ptsIntelligence Index

Artificial Analysis results from Oct 19, 2026: the best variant of each model family, with no estimated values. The index combines many tests and does not say which model will work on your task: that is what TestLAB is for. Source: Artificial Analysis

Methods, data and evals