


Task factories for useful intelligence. Research and design lab building expert-authored tasks and pixel-perfect environments to train and evaluate models.


Claude Fable 5.2 lands 2nd on GDPx legal, behind Muse Wonder Full, up 4.6 points on Fable 5.1.
Fable 5.2 debuts at the top of the GDPx leaderboard scoring 16.4%. (1/7)🧵

Dedicated website: gdpx.semianalysis.com/?domai… Blog post: newsletter.semianalysis.com/…

For GDPx, Chakra Labs contributed 20 legal tasks developed with senior-level attorneys across eight practice areas, from M&A and venture capital to intellectual property and dispute resolution. The rubrics accept supported alternatives wherever reasonable lawyers could disagree. They give greater weight to the reasoning behind a conclusion rather than just the end outcome. Importantly, weights scale with the consequences for the client, so missing the central legal issue costs more than a formatting error.

Great legal work depends on judgment, and two experienced lawyers can reach different conclusions on the same question and both be right. That makes legal benchmarks hard to design, since a rubric that credits only one answer penalizes the other.

Chakra Labs’ mission is to maximize access to useful intelligence. We're excited to partner with @SemiAnalysis_ on GDPx to measure how top models perform under realistic scenarios in legal domains.
Today we’re releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains. GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)🧵

Fantastic day at @tryterac’s inaugural RL Summit - great mix of eng, research, and applied folks all in one room. Thanks to @zacbakerr and @RyanMWexler for putting together such a sharp group, and for letting me share what we’re seeing at @chakra_ai. Excited to return next year!





To get started, check out our docs and code examples below. docs.trydojo.ai/integrations…

Ex: access our Amazon clone using familiar Gymnasium style APIs.

Now, you can access any Dojo clone using the same APIs leveraged by OpenEnv.

OpenEnv is a community framework developed by @huggingface, @PyTorch, and others for creating, deploying, and using isolated execution environments for agentic RL training.
