22 new models last week. Your tools heard about none of them.

Your coding agent doesn't know about last week's models.

Twenty-two of them shipped in the last seven days. Microsoft released five, Qwen four, Anthropic two, and a company called InclusionAI two more. I counted them on Monday. That's a real part of my job now.

It's hard to keep up with new models. It's harder to keep your tools up to date.

Here's a normal week for a developer with AI tools in 2026. A new model comes out. You open the pricing page. You check if your coding agent supports it. You check the context window, the price per million tokens, whether it can call tools. Then you update your config. Next week you do it again. And your agent, the thing that was supposed to save you time, waits while you do its homework.

Every week there's a model that changes everything. By Friday it changes back.

When we started Claudish, our open-source CLI that runs Claude Code with any model, we kept the model list by hand. A config file, updated by me, when I had time. You can guess how that went. Users opened issues about models I'd never heard of. Once I added a model a week after everyone stopped talking about it.

The config broke again.

So we tried a spreadsheet. The spreadsheet died in three weeks.

"Another model list? Really, Jack?" — you'll ask me. And it's a great question.

There are good model lists already. But a list is a snapshot. A person writes it, on one day, for other people to read. An agent can't use a snapshot. An agent needs to ask a question. Give me the cheapest model with tool calling and at least 200k context. A blog post can't answer that.

So we built a model catalog server, models.madappgang.com, and pointed our own tools at it first.

The idea is simple. Twenty collectors run twice a day. Thirteen of them talk to vendor APIs directly — Anthropic, OpenAI, Google, OpenRouter, Together, Mistral, DeepSeek, Fireworks, OpenCode Zen, X.AI, Moonshot, Zhipu, and DashScope. Seven more read the official pricing pages. A merger takes all of it, removes duplicates, and scores every value.

Every value carries one of five confidence tiers. Data from a vendor's own API is api_official. Data from an aggregator is aggregator_reported. A scraped price we couldn't verify is scrape_unverified. When two sources disagree, the field doesn't get a coin flip. It goes to a human.

That's the part I care about most. The reviewer can update a value, fill an empty one, skip it, or escalate it. It never invents a value to make a row look complete. Every change needs source evidence. I get model recommendations from LinkedIn, from Twitter, from my barber. He's actually pretty good. The catalog isn't allowed to listen to him.

Think of it as a flight timetable for models. Airlines publish flights. The timetable collects them, checks them, and keeps them current. Your travel agent reads the timetable, not the airline ads. Our tools read the catalog the same way.

Today the catalog holds 923 current models from 68 providers. That's 1,334 rows if you count the deprecated ones. It grows by about 60 models a month. When Anthropic shipped claude-opus-5 and claude-opus-5-fast on July 24, both were in the catalog the same day.

For agents there's a free public API. No key, no signup:

curl "https://us-central1-claudish-6da10.cloudfunctions.net/queryModels?minContext=200000&maxPriceInput=3"

An agent asks for the cheapest model with tools and a 200k window. It gets a real answer, with pricing and capabilities, updated this morning. You can filter by provider, search, or status, and cap the result count. There's a slim mode for CLI routing — Claudish uses it to resolve model names — and a recommended mode for ranked picks.

For humans there's the portal, the Claudish Models Portal. The timeline view shows what shipped and when. Compare puts two models side by side. The agents view shows which models work in which coding agents. Defaults and recommended give you a starting point. The subscriptions view answers a question I get every week. Which subscription actually gives you which models.

We built this for ourselves first. Claudish stopped hardcoding models, and my Monday counting ritual is over.

If your agent still reads blog posts to pick its models, give it the API instead. And if you find a model we missed, tell me. The collectors run twice a day, but the internet moves faster.