Skip to content
AI Dev Toolkit.
Esc
  • Loading…

Methodology

How prices are collected, how tokens are counted, how tools and fixes are tested, and how mistakes get corrected.

Last updated

Model prices and limits

Model data comes from public sources that don’t require an account:

  • The OpenRouter models API, which lists hundreds of models with their prices and context lengths.
  • The open-source LiteLLM model price and context-window file, maintained by its community.
  • Each provider’s official pricing and model documentation pages, used to check and correct the above.

A scheduled job runs every day. It downloads the sources, normalises them into one format, and checks every record against a strict schema. If anything changed, the update is published and recorded in the history. Corrections that the automated sources get wrong go into a manually maintained overrides file, which always wins.

If a source is unavailable or returns data that fails validation, the job keeps the last good data instead of publishing something empty or broken. Every price on the site shows the date it was last checked and links to where it came from.

Token counts

Models read text as tokens. Each model family splits text differently, so the same paragraph can be a different number of tokens for different models.

  • Exact counts are used where the tokenizer is public. For example, OpenAI’s tokenizers and the open-weight DeepSeek, Mistral and Qwen tokenizers run directly in your browser.
  • Estimates are used where the tokenizer isn’t public, such as recent Claude and Gemini models. They’re always labelled as estimates. Where the provider offers a free token-counting endpoint, you can get an exact count with your own API key.

Counts are for the text you enter. A real API request also uses tokens for chat formatting, system prompts and tool definitions, so your bill can be slightly higher than the count for the text alone.

Calculations

Cost calculators multiply token counts by published per-million-token prices, showing input, output and cached prices separately where providers price them differently. Hardware calculators show their formula and assumptions on the page, so you can check the working. Results don’t include taxes, discounts or free tiers unless the tool says so.

How tools are tested

The logic behind every tool is a set of small functions with automated tests. Tokenizers are checked against reference outputs for a fixed set of sample texts. Calculators are checked against hand-worked examples. Format converters and generators are checked against known-good files. The full test suite and a browser test of every page run before each release, and the daily price update is only published if the tests pass.

Error fixes

Every message in the error database is checked against the tool’s own error reference and the text inside the current build, so the wording matches what you see. Fixes follow the vendor’s documentation, rewritten with more context; steps that were run on a real machine are marked as tested. Each page states the tool version and the date it was last verified. Fixes aren’t copied from other sites.

Writing and AI assistance

Guides and explanations are written with the help of AI assistants, then reviewed and published by Tahir Nazir, as you’d expect on a site about AI developer tools. AI assistants also help build and test the tools. Every guide is fact-checked against primary sources before it’s published: facts come from the providers’ own documentation or from measurements, each guide lists its sources and says what was checked by hand, and numbers such as prices are computed from the same daily data as the tools, so they can’t drift out of date.

Prices are the exception: they update automatically every day from the sources above. Each update must pass the schema checks and the sanity checks described in model prices, and the history is spot-checked by hand. When an automated value is wrong, a manual correction is added that overrides it on every future update.

Corrections

If you spot a mistake, email hello@its-tahir.com with a link to the correct information (see what to include). Confirmed errors are fixed promptly and the page’s “last updated” date changes. Significant corrections are noted on the page.

Independence

No provider pays for placement, rankings or reviews. Comparisons are ordered by the numbers or by the sort you choose. If the site ever carries sponsored content, it will be clearly labelled as sponsored.

Keep exploring