Find the AI model your project really needs — not the one everyone says is best.
Describe your project in plain words. XFMS discovers what it truly requires, ranks every model against that on evidence, and proves the top picks on your own work. Discovery is free.
Free discovery. Ongoing monitoring with XF Model Watch from $19 a month. Optional live tests run on your own OpenRouter account and cost a few cents.
Works inside Claude Code, Cursor and other assistants that support MCP plug-ins.
Every platform has a model recommendation. None of them did the work for your job.
Ask around and you get a confident answer. It just isn't about you.
Popularity is not fit
The model everyone names was not measured on your job.
The platform picks its own
Whoever hosts the recommendation sells the answer.
The router won't show its work
A black box routes your call and never says why.
Guesswork is cheap. Being wrong on the model you build on is not.
Three steps, in the assistant you already use.
No benchmarks to read and no ML background required — if you can describe the job, XFMS can rank it.
1. Discovery: what your project really needs
Your assistant asks a few plain questions. XFMS turns your answers into the project's real requirements, the ones a leaderboard never asks about.
2. Ranked on evidence
Every model is checked against independent test results, prices and features. The data is refreshed weekly.
3. Proven on your own work
The top picks do your real tasks side by side. You get a winner and the reasons.
Finding it is free. Discovery and ranking cost you nothing. Only the side-by-side test costs anything, a few cents on your own account, and only after you say go.
Catching the models that break the format you need.
From our daily tests on 15 October 2026. Three invoices is a small sample, so read this as how the method works, not as proof of broad reliability. The same model won all five times we ran this job.
The job
An agency automating bookkeeping for clients needs the invoice number, date, total and vendor pulled out of each invoice as JSON, about 2,000 a month. The cheapest model that gets every field right is fine.
What the job really needs
Not "the smartest model". It needs these, read back before anything ran:
Must return valid JSON every time. Every field must match the invoice. Cost decides among those that pass.The candidate models
Five models ranked on evidence for this job, plus OpenRouter's own automatic picker, all given the same three invoices.
Ling 3.0 FlashMistral NemoLlama 3.1 8B InstructLing 3.0 Flash VLgpt-oss-20bOpenRouter's pickerThe evidence
Every answer was checked by code for valid JSON and by a grader for every field. Llama 3.1 8B wrapped its JSON in chat on all three invoices and was ruled out. Mistral Nemo, the cheapest, broke the JSON once and was held for a second look. Failures are quoted word for word on the results page.
$0.00discovery and ranking$0.0349the whole test3 of 3invoices the winner got fully right3 of 3invoices where Llama 3.1 8B wrapped its JSON in chatThe decision and the reason
Ling 3.0 Flash passed every must-have and cost the least among the models that did, which is exactly how this job said to choose. OpenRouter's picker also passed, but came last of the four that did.
Winner: Ling 3.0 Flash · inclusionai/ling-3.0-flash
That is the same discovery you get free, on your job.
more than 200 ways models differ, across more than 450 models.
Leaderboards compare models on a handful of scores. Model Source tracks every published difference, refreshed weekly, and discovery uses whichever ones your job actually depends on.
Some unusual abilities are known only from the provider's own claims; we label those as lower trust.
What people ask before they trust a pick.
Why not just ask ChatGPT?
A chat assistant answers from memory: the models it has seen praised most. It hasn't measured them on your job, and it may not know this month's models or prices. XFMS starts from your job's real requirements, scores models against collected test results, and shows the evidence for each pick. Then, if you want, it proves the best ones on your own work.
Is my data and my key safe?
Your own assistant holds the conversation. XFMS receives only the answers it needs to score your requirements, and keeps no conversations: your progress travels in a sealed note your assistant carries, which expires after an hour. We log the one-line description you start with and an anonymous ID, to count usage. Your OpenRouter key is used only for a test you approve, for that run, and is never stored.
Why should I trust the rankings?
Every week we refresh prices, features and test results for every model on OpenRouter, from independent sources. For each of your requirements, a model is either measured, with how sure we are shown, judged from related tests, or has no evidence. No evidence is shown as unknown, never as good. Every pick shows the evidence behind it.
What does it cost?
Discovery and ranking are free. A side-by-side test runs the top models on your own tasks through your own OpenRouter account, usually a few cents. You see the likely cost and the most it can cost before anything runs, and nothing runs until you say go.
Try it free in the assistant you already use.
You need Claude Code, Cursor or another assistant that supports MCP plug-ins. Add one address, then ask your assistant to help you pick a model. Your first three discoveries need no sign-up and no key.
Claude Code
Paste this in your terminal:
claude mcp add --transport http xfms https://xfms.xpansion.dev/mcp/
Cursor and other assistants
Add an MCP server with this address:
{ "mcpServers": { "xfms": { "url": "https://xfms.xpansion.dev/mcp/" } } }
Then ask
Paste this to try it, then swap in your own job:
Help me pick an AI model for pulling the invoice number, date, total and vendor out of invoices as JSON, about 2,000 a month. The cheapest one that gets every field right.
For a side-by-side test
Add your OpenRouter key as the X-OpenRouter-Key header, only when you want XFMS to run the top models on your work.
Watch my app.
Free finds the right model today. XF Model Watch keeps it right. New models and price cuts land every week. Save your project once, and we re-check it weekly on your own tasks. When a cheaper or better model passes all your checks, you get an email with what switching would save, and a warning before your model is retired. You decide whether to switch.
Your comparisons are on us while you subscribe. Two months free when you pay yearly. 14-day money-back promise.
Not ready yet? Get the free email: we test new models as they arrive and tell you what they're good for, plus a weekly summary of what got cheaper or better for jobs like yours.
Building this for a whole business?
Model Source is one part of Xpansion. The same discovery-first method maps what a business actually needs and decides what to build, before anything is built.