ELO SI
Open Brain
An AI that learns what it has never seen.
Research, early.
What it is.
Open Brain learns worlds it has never seen. Shown a few examples of a small world with hidden rules, it works the rules out well enough to say what happens next, or to use it.
Then it is tested on worlds no mind has seen before. That is the skill we care about: learning something new, quickly, from very little.
The exam.
We measure Open Brain on the Golden Learning Level, our own exam. Every world in it is sealed: made fresh from a new seed, so nothing in it can have been seen or memorised. A mind gets a few examples, then answers.
Prediction. A few steps of a hidden process; say what comes next.
Working an app. An app it has never seen; use it to reach a goal.
A score is quoted only when two fresh seeds agree, and every quoted run is replayed. Claude Opus 5.5 sits the same exam, run the same way.
Where it stands.
- Prediction
- Level with Claude Opus 5.5
- Working an app it has never seen
- Level with Claude Opus 5.5
- Reading English
- Learning now
Our exam, run by our lab. It is not an independent benchmark. Claude and Opus are trademarks of Anthropic.
Our vision for 2027
An AI that does research.
By 2027 we want an AI that doesn't only answer, but finds things out: one that designs its own experiments, runs them, checks the results and improves itself, with people deciding what it may change. These are goals, in order. Only the first is done.
- 1
Learn new worlds from a few examples. Done
Level with Claude Opus 5.5 on our first two exams, Prediction and working an app it has never seen.
- 2
Read and write. Now
English it understands, then writing of its own.
- 3
Every field. Next
Reading, writing, reasoning, mathematics and code, each on sealed exams against the best models of the day.
- 4
Proofs a computer checks. Next
Mathematics where every result is checked by a proof assistant, so a claim is either proved or it doesn't count.
- 5
Its own researcher. Next
The lab's experiments designed, run and judged by Open Brain itself, with a person approving every change it makes to itself.
Our standard.
Measured, not claimed.
Nothing about Open Brain is said until a sealed exam shows it.
Against the best.
Every result sits beside the strongest models of the day, on the same exam.
People in control.
Nothing it changes about itself ships without a person's approval.