Best Fit

Describe what you need.
Get the model that fits.

Type it in plain English. We rank the models on real test results.

Best coding modelPrivate chat on my RTX 4060, 32GB RAMRead scanned invoices into JSONSort emails as cheaply as possibleFast transcription with speaker labelsA natural voice for audiobooks

Use Best Fit inside Claude, ChatGPT, Codex or Cursor →

Ask Best Fit from your AI appClaude, Codex, Cursor and other assistants can use Best Fit directly: free, no key.

https://bestfit.dev/mcp

Add to Cursor Add to VS Code VS Code Insiders

In a terminal:

claude mcp add --transport http bestfit https://bestfit.dev/mcp

In a terminal:

codex mcp add bestfit --url https://bestfit.dev/mcp

In a terminal:

gemini mcp add --transport http bestfit https://bestfit.dev/mcp

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "bestfit": {
      "serverUrl": "https://bestfit.dev/mcp"
    }
  }
}

Settings → Connectors → Add custom connector, and paste:

https://bestfit.dev/mcp

Settings → Apps & Connectors → Advanced → Developer mode → Create, and paste:

https://bestfit.dev/mcp

Use the address above in the app’s MCP settings; an app that only runs local servers can bridge to it:

npx -y mcp-remote https://bestfit.dev/mcp

Full setup guide →

How it works

How Best Fit works

Tell it the job. It ranks every current AI model on real, independent test results for that job, with what it costs to run.

You describe the job“cheap, good at code”

We pick the teststhat measure that job

You get a rankingfrom real results

Real results, not opinions

Every score comes from the organisation that ran the test. Makers’ claims never rank.

Ranked for your job

Only the tests that measure what you asked for count.

The cost you’ll really pay

Running cost per task, thinking included, not just the token price.

Fresh every day

Every board is re-read daily, so new models show up within a day.

Honest about doubt

Estimates are marked, margins of error drawn, ties called ties.

Works in your AI app

Claude, ChatGPT, Codex and others can ask Best Fit directly.

—models

—tests

—independent testers

—test results

Reading a chart

A dot is a modelColour is the maker.

Bigger is fasterDot size is output speed.

A line joins settingsOne model’s thinking levels.

A bar is the marginOverlapping bars: no clear winner.

Dashed is estimatedNever mixed with measured.

Green line is good enoughBelow it, a model does not rank.

Tap ? on any chart for its own key.

Drag the triangle

	Model	Fit	Quality	Cost per task	Speed

An example: move the dot and the table re-ranks.

What we rank

Language modelsChat, coding, agents, maths, 16 languages.

Image generation and editingBlind votes between two pictures.

VideoFrom text or an image, with and without sound.

Text to speechBlind votes between voices.

Speech to textWords right on real recordings.

Embedding modelsA short list in rough bands, not a ranking.

Not ranked yet3D generators, music and real-time voice agents. Best Fit says so and suggests the nearest thing it does rank.

Use Best Fit inside your AI app

Bring Best Fit to your tools

Best Fit is an MCP server: Claude, ChatGPT, Codex, Cursor and other assistants can ask it which model fits a job, using today’s data.

Free, no key, read-only. Tools: find_model and get_model.

On claude.ai, the desktop app or the phone app: Settings → Connectors → Add custom connector. Name it Best Fit and paste this address. No sign-in is needed.

https://bestfit.dev/mcp

In a terminal:

claude mcp add --transport http bestfit https://bestfit.dev/mcp

Where your plan allows your own connectors: Settings → Apps & Connectors → Advanced → Developer mode, then Create, paste the address and choose no authentication.

https://bestfit.dev/mcp

In a terminal:

codex mcp add bestfit --url https://bestfit.dev/mcp

or add to ~/.codex/config.toml:

[mcp_servers.bestfit]
url = "https://bestfit.dev/mcp"

Add to Cursor or put this in ~/.cursor/mcp.json:

{
  "mcpServers": {
    "bestfit": {
      "url": "https://bestfit.dev/mcp"
    }
  }
}

Add to VS Code or put this in .vscode/mcp.json:

{
  "servers": {
    "bestfit": {
      "type": "http",
      "url": "https://bestfit.dev/mcp"
    }
  }
}

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "bestfit": {
      "httpUrl": "https://bestfit.dev/mcp"
    }
  }
}

Any app that speaks MCP over HTTP can use the address above: no key, read-only. An app that only runs local servers can bridge to it:

npx -y mcp-remote https://bestfit.dev/mcp

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "bestfit": {
      "serverUrl": "https://bestfit.dev/mcp"
    }
  }
}

The detail

What happens to your question

An AI reads it once. Jev, a small model from TypeSafe, works out the kind of model, which tests matter, how good a model the job needs, and how much quality, cost and speed matter to you.
Code does the sums. Real results on those tests become Quality; models that can’t do the job, or fall below the level it needs, drop out. Quality, cost and speed, balanced as the triangle shows, make Fit.
Drag the triangle. Move the dot towards Best, Cheapest or Fastest and the table re-ranks at once.
Good enough is enough. On a simple job, extra points on hard tests never win a model many times dearer.
One vote per skill. Boards that measure the same thing share a vote.
How close is the closest test? A line under each answer says how directly our tests measure your job.

Where the numbers come from

Each result is read from the tester’s own site or files. Testers: .

Prices are the makers’ own, as Artificial Analysis lists them, or OpenRouter’s where they list none. Costs are the testers’ own; if a price has changed since, the cost is re-priced and marked.

Specialist tests

Medicine: patient records, diagnosis codes, visit notes, drug discovery, clinical images.
Law: contracts, case-law research and case documents, marked on a lawyer’s rubric.
Finance: company filings, Excel models, tax research, a month-end close.
Documents: invoices, receipts, charts and tables.
Languages: the same 400 exam questions in 16 languages.
3D and design: floor plans, 3D scenes, drawings, websites, charts and games.
Trade arenas are ordinary users’ votes: preference, not correctness.

How we keep it right

Questions asked here and through the MCP are kept, with nothing that says who asked, so we can see what people need.
Daily: every board re-read and checked against the tester’s files; anything odd is held back until checked.
Weekly: our numbers compared with every tester’s web page, and a few charts checked by eye.
The method is checked against three independent AI judges on 87 questions whenever it changes.
Shared links freeze the answer as it was that day.

Arena ratings are taken from lmarena-ai/leaderboard-dataset, used under CC BY 4.0.
The language boards are Cohere Labs’ Global MMLU Lite, run on Kaggle Benchmarks and used under Apache 2.0 (Singh et al., Global MMLU, 2024), as are the FACTS boards.

Best Fit

Describe what you need.
Get the model that fits.

New models

How Best Fit works

Bring Best Fit to your tools

Get in touch

Describe what you need.Get the model that fits.

New models

How Best Fit works

Bring Best Fit to your tools

Get in touch

Describe what you need.
Get the model that fits.