April 15, 2026
The results are in: Meet our Cohort 0 winners
The first-ever Arena challenge is over. Here’s the full recap of Cohort 0’s Grounded Reasoning challenge on OfficeQA — and the teams that took home the prizes.
The results are in: Meet our Cohort 0 winners
The first-ever Arena challenge is officially over. Before we kick off the next phase, here’s the full recap of what Cohort 0 achieved.
Introducing Challenge 0: Grounded Reasoning
Over the past three weeks, a cohort of 147 of the best AI builders and researchers from a pool of 1,200+ applicants took on Grounded Reasoning, our inaugural challenge based on the OfficeQA benchmark by Databricks. Across 9 events, 3,000+ RSVPs, and 495 in-person attendees, Cohort 0 didn't just build — they showed up. This is why Arena isn’t just a hackathon, but a network and community of open-source builders.
At the center of it all was a single goal: to build AI agents capable of extracting precise numerical answers from U.S. Treasury bulletins spanning 1939 to 2025. Each agent receives a financial question and an environment containing the bulletins, reads the documents, reasons over the financial tables within, and returns an exact answer. This turns decades of messy government records into reliable data through grounded reasoning.
To compete, participants picked a pre-built coding agent — Codex, OpenHands, goose, or OpenCode — and tuned it through prompts, skills, tools, and configs. For the first time, Arena participants also actively experimented with AI self-improvement tools — specifically the alpha version of EvoSkill — to level up their builds by combining human insight with AI evolution.
Introducing the Winners of Cohort 0
1st Place — The Big Q ($16K in prizes)
With Anmol Sharma, Siddharth Bhat, and Jonathan Kim
The Big Q built an open coding agent that uses RCI (Recursive Criticism and Improvement) to answer complex financial questions. This method involves a primary agent generating an initial answer and a secondary agent that critiques it against source documents, refining each output in a continuous feedback loop until correct. With each successful session the team was able to teach the agent how to read tricky table layouts and eventually create a scalable framework for other fields such as healthcare.
2nd Place — Dolores Research ($12K in prizes)
Solo team with Leon Liu
3rd Place — PURSUIT ($6.5K in prizes)
Solo team with Ajinkya Mulay
By building a Sanskrit-inspired reasoning framework called Karaka Router, Ajinkya was able to use his agent to extract 89K pages and 26M numbers worth of data from the U.S. Treasury Bulletin corpus. In doing so he was able to push past the 70% accuracy limit reached by Claude Opus 4.5.
Honorable Mentions
4th–10th Place ($1.5K each in prizes)
These teams held their own against the best builders and researchers in Cohort 0 and proved that they belong in the top 10.
- 4th Place — GroundWire, Cat Varnell
- 5th Place — RETRO, Robert Amanfu
- 6th Place — Zero Node, Jwalin Shah, Tarun Reddi, Sanjay Sai
- 7th Place — Zeno AI, Shubham Agrawal
- 8th Place — WINNERS, William Chen
- 9th Place — Tea-tiok, Vikranth Reddimasu
- 10th Place — Kurukshetra's, Dhara Goyal
Finally, shoutout to all the teams who participated and made the competition what it was. We thank you for pushing the competition forward and raising the bar for everyone in Cohort 0!
In Conclusion
Arena Cohort 0 saw over 1.2k applications and 147 accepted participants. Every accepted participant received $400 in free compute from our friends and Arena sponsors OpenRouter, Daytona, and Dedalus Labs to help them get started.
This challenge wouldn't have been possible without our sponsors and partners:
Sentient Labs ($11K in cash prizes), Fireworks AI ($10K in credits), Daytona ($18K in credits), Yotta Labs ($6K in credits), and OpenRouter ($13K in participant development credits).
And thank you to everyone who supported Arena Cohort 0 — Franklin Templeton, Founders Fund, Pantera Capital, OpenHands, Dedalus Labs, and alphaXiv.
We also want to recognize the community events that brought Arena together — from the opening ceremony and Poker Night to our NVIDIA GTC Ice Cream Cart and Demo Day that packed the house.
Thank you to every single participant. Cohort 0 was everything we hoped it would be and more.
What's Next
This was just the beginning.
The Arena is coming back — with new challenges, bigger prize pools, and room for many more builders. Truly open-source.
If you missed Cohort 0, your chance is coming. Stay tuned.
- Check out the Sentient Arena website
- Watch the replay of our Demo Day livestream on YouTube
- Catch up with our Arena YouTube playlist
