ModelCensusopen-source ai reliability harness
Blog30 Apr 2026strategypanelcrystal-ball

Your benchmark is stale before it publishes

Our panel spans three years of releases. Half of it shipped in one year.

The signal most evaluation has not absorbed: the gap between a model and its successor is now measured in months, and nearly every eval is built on the assumption that the thing you measured will still be there when you publish.

2023 releases on our panel2202462025102026 (to Feb)2
Twenty models by release year. Half shipped in a single year.

The forces

Competitive cadence is the obvious one. The less obvious one is that a model id is not a fixed artefact — behind it sit providers whose quantization and configuration differ, and both the weights and the routing change without the id changing.

So the thing you measured can move without anyone announcing a release at all.

Old way, new way

Obsolete: benchmark-a-generation as the unit of work. By the time it publishes it describes something that has been superseded twice.

Valuable: a fixed set of cases, conditions and detectors versioned together, so a new model is one more loop rather than a new experiment.

The timeline

Immediate: teams re-run benchmarks more often and notice the numbers move without explanation.

Near-term: those movements get attributed to model quality rather than to routing, because almost nobody records which provider actually served a request.

Dominant: provenance becomes the differentiator. Not the score — the ability to say what produced it.

What to do now

Record the model actually served, not the one requested

One field. It costs nothing and it is the difference between a contradiction and a new observation when you re-run in six months.

Why: Without it, a changed number looks like your eval is unstable rather than like the endpoint moved.

A leaderboard is a claim with a shelf life. A failure taxonomy is a claim about the shape of the problem.

How would you know if the model behind your API id changed tomorrow?

Every figure here describes something measured and committed. See the measurements · read the method