The models behind AI Storming

The current roster

Every session draws from these six. You choose which ones join, which tier each runs at, and which of them moderates. The seat is earned on what the model contributes to an argument — not on where it sits on a leaderboard.

Four of the six now run one model across both tiers, with the fast/thinking switch changing how long that model reasons rather than which model answers. That is a change in how the frontier labs ship: depth has become a dial rather than a separate product.

How much a benchmark number is worth

Published scores are useful. They are also softer than they look, and the softness is not hidden — you can see it in the four examples already on this page.

The same benchmark, two variants, forty points apart

"SWE-bench" names a family, not a test. A model reported around the mid-nineties on SWE-bench Verified can land in the sixties on SWE-bench Pro. Neither number is dishonest. But a headline that says only "SWE-bench" has told you almost nothing.

The conditions belong to whoever is reporting

The clearest recent case: a 38.3% ARC-AGI-3 result was published for a frontier model with two provider-specific API features switched on. Run through the ARC Prize project's own harness, the same model scored 7.8%. The gap is not fraud — it is the difference between "what this model can do with our scaffolding" and "what this model does on neutral ground." Almost no headline distinguishes them.

Labs publish the tests they win

One of the models on this panel launched with a strong set of agentic and coding results and simply no GPQA Diamond figure. That is a normal commercial decision, and it means an absence on a launch page is information too — just not the kind you can put in a table.

Some numbers have not been reproduced yet

A vendor's first-party figures are a claim until someone outside the vendor reproduces them. Open-weight releases are the honourable exception: when the weights ship, anyone with the hardware can check.

Every one of these numbers describes a single model, working alone, on a task with a known answer, under conditions its own vendor selected. Your question has none of those properties.

How a model earns a seat — and how it loses one

Frontier releases have been landing every few weeks. The roster is reviewed against them continuously, and this page moves when the roster does.

Three things decide a swap. Independence comes first: one seat per lab, because the panel's value collapses if two seats share a training lineage. Contribution to an argument comes second, and it is not the same as a benchmark score — a model that reasons beautifully but concedes instantly is worth less to a debate than one that is merely stubborn and specific. Behaviour under our own conditions comes third: long shared context, several rounds, a moderator, and a hard requirement to address what the previous speaker actually said.

Before any swap ships, the new model is checked against its provider's API documentation and then against the live API. That second step is not ceremony. In the August 2026 review, four of the five providers we re-examined had documentation that contradicted their own API on at least one point that mattered — a parameter documented as supported that returned an error, a setting documented as unavailable that worked, a default that was not the documented default. A roster kept current from release notes alone would be quietly wrong.

The panel is a composition, not a ranking. The aim is six models that fail differently — not the six highest scores on any one list.

Frequently asked questions