Methodology

AI Model Performance: what has been published today about 15 physical AI foundational models, based on the six main criteria. Where nothing has been published, the cell says so. It is a draft standard, put up for discussion.

The six axes

Axis	Question	Measured as
Success	Does it do the trained task?	% of trials, trained setup
Robustness	Does it survive the cell drifting?	% under the perturbation set
Instruction	Does rewording the instruction break it?	% under paraphrase
Recovery	Does it fix its own mistakes?	% of injected failures recovered
Speed	Is it as fast as a trained person?	human cycle time ÷ robot cycle time
Interventions	How often does a person step in?	per 1,000 cycles

Evidence grades

Every published value carries a grade for how much of the method behind it is public:

A — Independent, with logs.
B — Paper or report with method (rubric or error bars).
C — Claim; method unpublished.

On the charts, a grade-C value is drawn hatched (striped): a number was claimed, but the method that produced it was not published.

Reading the marks

Real robot results and simulation results are kept apart everywhere: filled marks are real, hollow marks are simulation. A simulation score is not a real-robot score.
~ before a value means it was read from a published chart rather than stated as a number.
Owed means the vendor or authors have not published a result for that axis. The reason on record is shown with it.
A 95% interval, shown on request on the evidence strip, is the range the true success rate plausibly falls in given how many trials it rests on (a Wilson interval), and appears only where the trial count is published. Nine successes in ten trials spans about 60–98%; 900 in 1,000 spans about 88–92%.

Sources

Every value was checked against its primary source on 23 Oct 2026. Each one links to that source; where the browser supports it, the link scrolls to and highlights the figure. Sim-versus-real comparisons pair the same policy on the same task where such pairs exist, and are marked when they compare different tasks.

Download the data

One row per model and axis, published or owed, with value, setting, grade, trial count, context, note and source:CSV ·JSON. An MCP server, so AI agents can query it directly, is planned.

Connect an AI agent (MCP)

The scorecard is also an MCP server, so an AI assistant can look values up, compare models and cite sources directly. Read-only, free, no account or key.

Any MCP client

Add a remote MCP server. Transport: Streamable HTTP. Authentication: none.

Server URL

https://mcp.b2atoms.com

Claude (claude.ai and Claude Desktop)

Settings → Connectors → Add custom connector, then fill in:

Name

B2A AI Model Performance

URL

https://mcp.b2atoms.com

Claude Code

In a terminal:

claude mcp add --transport http b2a https://mcp.b2atoms.com

ChatGPT

Settings → Apps & Connectors → Advanced settings → turn on Developer mode → Create. Authentication: No authentication.

Name

B2A AI Model Performance

URL

https://mcp.b2atoms.com

Cursor

Add to ~/.cursor/mcp.json (or merge into the mcpServers you have):

{
  "mcpServers": {
    "b2a": {
      "url": "https://mcp.b2atoms.com"
    }
  }
}

VS Code

Add to .vscode/mcp.json in your project:

{
  "servers": {
    "b2a": {
      "type": "http",
      "url": "https://mcp.b2atoms.com"
    }
  }
}

Tools: list_models, get_model, compare_models,rank_by_axis and get_methodology. Answers carry the same grades, settings and sources as this page.

How to cite

Bits 2 Atoms, "AI Model Performance (policy scorecard)", b2atoms.com/data/policy-scorecard/, values checked 23 Oct 2026.

Corrections and new results

If a value is wrong, out of date, or a result has since been published for an owed axis, submit it with its source. Nothing submitted appears here until it has been checked against the primary source, the same as every value already on the page.