Frontier Rating today
Every flagship model, one rating, and an honest guess at what comes next.
Labs publish benchmark scores in a different mix every time, so raw percentages are not comparable — and they run out, because every benchmark eventually saturates. AGI Frontier fits one unbounded Frontier Rating per model from the numbers each lab reported itself: not an average, but the latent ability of a Rasch model, where 400 points is ten times the odds of solving an average item of the basket. The bounded Frontier Index (0–100) is the same scale read differently. On top of it sit levels and their crossing dates, the pace eras, and a release forecast per lab whose window is as wide as its own back-tested track record says it must be. Everything on this page is recomputed in your browser from one auditable file, back to the first GPT.
The frontier
Loading the frontier…
How the forecast did
The scrubber is in the past, so the forecast made on that date can be scored against what actually shipped. Green means the release landed inside the 68 % window, amber inside 90 %, red outside it.
Stages
One column of time, read downwards: the levels the frontier has not reached yet with their windows, then NOW and the pace it is travelling at, then the eras behind it and the levels it has already passed. Every level is fitted from the benchmark difficulties — none of it is written by hand.
Release watch
Per lab, as of the scrubbed date: how long since the last flagship, and the log-normal cadence model's probability that the next one lands inside 30 and 90 days.
Rankings
Tiers
The newest released flagship per lab, ranked by Frontier Rating. Select a row to open its audit.
| # | Model | Rating | Index | Benchmarks used | Released |
|---|
What changed
The tail of the append-only audit log — every edit the worker or a human made.
Researcher
The published dataset is produced by an automated researcher through OpenRouter, and it is only promoted once it reproduces the frozen, human-curated gold set closely enough. Its schedule, its evaluation and its budget are published here for the same reason every score carries a quote.
Method
A flagship is the most capable tier a lab offers at launch; mid and small tiers are recorded too, feed the fit and the family bands, but never the frontier line. Only numbers a lab published itself count as official, and every score carries a URL and a verbatim quote that a worker re-checks. Averaging raw percentages would reward a model for reporting only easy benchmarks, so instead a one-parameter logistic (Rasch) model is fitted jointly over every official score: logit(score) = θmodel − δbenchmark. The Frontier Rating is 1000 + 173.72·θ — unbounded, so a saturating basket cannot flatten it; the Frontier Index is 100·σ(θ), the score a model would be expected to get on a benchmark of average difficulty. A hollow marker — matched by a "provisional" badge in the rankings and the release watch — means the release reported fewer than three index benchmarks, so its rating rests on too little evidence to form the frontier line. Release timing comes from a log-normal fit of each lab's inter-release intervals, shrunk toward the cross-lab prior, and widened by the σ scale the backtest earned. Scrubbing back in time re-runs the whole fit as of that date, which is how you check whether the forecast would have called what actually happened.
A rating is only as strong as the evidence under it, and the evidence has thinned. A model released in the first half of 2026 carries a mean of 5.5 index benchmarks; a model released since July carries 2.6, and some 2026 H2 flagships carry exactly one score — an LMArena Elo, which is a vote count rather than a test result. Two models a hundred rating points apart can therefore be statistically indistinguishable, so the rankings bracket rows the data cannot separate under a shared rank and report the evidence each rating rests on instead of an order alone.
| Released | Models | Index benchmarks each |
|---|---|---|
| 2025 H1 | 15 | 4.5 |
| 2025 H2 | 18 | 5.7 |
| 2026 H1 | 13 | 5.5 |
| 2026 H2 | 8 | 2.6 |
The dataset is also behind the field: on 8 Sep 2026, 23 of the LMArena top 45 models were not in it at all. The newest model a lab has here can lead its family without being that lab's newest model in the world, and a lab whose latest release has one score sits on a wider interval than one whose latest release has seven.
Data
The site loads exactly one file and computes the rest in your browser. Nothing else is fetched — fonts are self-hosted, there is no analytics and no third-party script.
- Last worker run
- —
- Last success
- —
- Pages polled
- —
- Extractor model
- —