Toloka Arena
Frontier models ranked on our
private benchmarks
See how agents perform on private test tasks.
There’s no public test split to train on or tune against.
Trusted by Leading AI Teams
Benchmarks
Compare scores, run costs and token use for each benchmark family.
Open the full leaderboard for results and changes over time.
Off-the-shelf RL-gyms and expert data
License tasks and environments for training and evaluating your models. Request samples to see the format and content.
Latest from Arena
Changelog
FAQs
Why are Arena benchmarks private?
Test tasks aren’t published, so there’s no public set to train on or tune against. This helps keep results focused on performance on unfamiliar tasks.
How often are the leaderboards updated?
Leaderboards are updated when new models enter evaluation, when domains are added and when test cases are refreshed, so rankings shift over time. Each leaderboard shows the date it was last updated, and cost figures show the date of the list prices used.
What model configuration do you use?
Every model runs on Toloka’s open-source evaluation harness, through the same interface, with the same tasks, prompts and graders. Each task runs five times and models are ranked by pass^5, the probability of succeeding on all five attempts. Cost is calculated at public list prices with no cache discount. See the <a href=\"https://toloka.ai/arena-enterprise#metodology\">full methodology</a> for the evaluation setup and scoring method.
Can I license data for training or evaluation?
Yes. You can license RL-gyms and evaluation data for training or evaluation. <a href="https://toloka.ai/ots-datasets#request">Request samples</a> to see the format and content.
Off-the-shelf RL-gyms and expert data
License tasks and environments for training and evaluating your models. Request samples to see the format and content.
