A self-updating catalog of 56 frontier models, with an autonomous research agent that writes new entries and a benchmark suite that scores models on launch day.
Comparing frontier models means hunting across four provider sites, a pricing page, and someone's thread on X.
AI Model Guide puts the specs, availability and distilled guidance for 56 models in one place —
and then goes a step further: the catalog largely maintains itself.
56Models catalogued across 5+ providers
0API keys needed to run the eval self-test
100%Static — every model page pre-rendered
The catalog
Each model is an MDX file with typed frontmatter — context window, modalities, pricing per million tokens,
availability, strengths, official docs.
Zod validates that frontmatter at build time, so malformed data fails the build rather than
shipping a broken page. The schema is the single source of truth, and adding a model is a matter of matching it.
Built with Next.js 16 on the App Router, statically generated, with Tailwind v4 for the
interface and static tweet embeds that need no X API at runtime.
The research agent
The interesting problem isn't rendering the catalog — it's keeping it current. A tracker can notice that a model ID
it has never seen just appeared, but turning that signal into a correct, sourced model page used to need a person.
That's the step the agent automates:
tracker → “grok-4.6 is new” → research agent → validated spec → scaffold → build → publish
The agent researches a candidate model and emits a spec that has to satisfy the same Zod schema as a
hand-written entry — so the validation gate applies equally to machine-written pages.
It runs in a dry mode with no key and no cost, which makes the pipeline testable without spending anything.
AMG Eval — scoring models on day one
Third-party leaderboards take days or weeks to score a new frontier model, which left the catalog able to describe a
launch but not compare it. So the project runs its own suite:
Deliberately small and honest rather than sweeping. It won't tell you which model is smartest —
it tells you how models compare on a fixed set of tasks under identical conditions.
Every number is reproducible: the suite self-tests with no API keys, scores a single model on
demand, or scores everything and writes the results back into the site.
The point is the methodology note as much as the score. A benchmark that overclaims is worse than no benchmark.
What I'd do next
Widen the eval suite while keeping the same honesty constraint — more tasks, same fixed conditions.
Track pricing over time rather than as a snapshot, since that's the field that goes stale fastest.