ORO

We’re excited to announce the next stage of our subnet: an automated post-training pipeline that will enable us to build the best product for AI shopping. It’s only been 50 days, but we’re now getting 20k high-quality trajectories per day that are rich training signals for online shopping tasks. Using the trajectories we have thus far, we saw an 18% → 42% climb on Qwen3-4B base using our post-training pipeline.

@evilsadsick all the way up

@oroagents only up for ORO

In the past 101 days of being live, we have judged 2.5M+ agent trajectories. These are agent decisions that allow us to understand where the state of today's AI shopping is at. We demonstrate the quality of this data with our own 4B post-train, with a climb from 18% to 42.7% on just 1 day's worth of data. You'll be able to visualize this very soon. We've also increased the payouts. The best builders on ORO are taking home up to $20k USD a day in on-chain rewards. Oh, also, we've been live for 101 days CONSECUTIVELY. 0 days of downtime. We've been scoring some big wins since our launch, and there's some even bigger ones in the pipeline. Stay tuned.

Great to have gone through EdgeBench. At ORO AI, we're working on some long-horizon agent evaluations as well that will change how benchmarking is performed for agentic commerce tasks. There's a reliability gap that SN15 is built to close. EdgeBench from the Bytedance team validates the market demand of continuous, real-world benchmarking infrastructure. Excited to share some more news in a bit...

Some big updates for the miners on ORO in the last 2 days. Daily on-chain rewards are now up to $20,000 USD per day for performing well on the ShoppingBench eval. That means in the past 2 days, these 2 miners have earned just under $40,000 cumulatively, by scoring a 71.7%: 🥇 Race #72 Winner: titan 🥇 Race #73 Winner: tower

@proofoftalk was an amazing experience, so glad we had the team there to be able to showcase our work and connect with builders and supporters of @bittensor

A few weeks ago we were on stage at Proof of Talk 2026, one of the biggest events in AI and Web3, held inside the Louvre. We had 15 minutes in front of a room of founders and investors to make three points: 1.⁠ ⁠AI is already shopping on your behalf. 2.⁠ ⁠There's no open benchmark to measure if it's doing that well, and; 3.⁠ ⁠We’re building the open, live arena that fixes both. Here's the talk:

@buildwithsamp Looks like we need to add baguettes to the list of things our agents will shop for! 🥖

τ + 🥖: a mini series where I breakdown concepts/subnets in Bittensor in the time it takes me to walk and get a baguette A walk through SN 15: Oro @oroagents Check them out at oroagents.com

A bite-sized explanation of what we are doing at ORO

Checkout this article and the rest of the articles written by our team - oroagents.com/blog

489 million requests in 24 hours, ORO miners and users had 0 impact. Reliability and availability are pivotal for us to run a consistent competition that brings the best miners to our subnet.

All of the data is open source already through operating the subnet. We're hoping the HF collection, the repo, and the paper can help you understand more deeply what our mission is.

Paper can be viewed here: arxiv.org/abs/2606.10064 HuggingFace collection: huggingface.co/collections/or… Repo: github.com/ORO-AI/shoppin…

We've talked a lot about how our efforts to train AI to shop will be entirely open source. Through Bittensor, we're committed to that ethos. We're excited to share our pre-print on arXiv, our code, our data and our entire post-training pipeline. Huge shoutout and thanks to @JarrodBarnes in helping us leverage this very valuable data. This is how AI is going to learn to shop.

Shardul @oroagents talking through Improving the open benchmark for what good shopping means at @proofoftalk @shardiban

Our very own @shardiban is going live @proofoftalk to talk about ORO and everything we have built for the subnet so far - x.com/i/spaces/1oJMv…

Crucible Labs and @oroagents in Paris for @proofoftalk

@Parth378 thank you for your support!

@oroagents Awesome stuff thanks for the update!

We’re excited to announce the next stage of our subnet: an automated post-training pipeline that will enable us to build the best product for AI shopping. It’s only been 50 days, but we’re now getting 20k high-quality trajectories per day that are rich training signals for online shopping tasks. Using the trajectories we have thus far, we saw an 18% → 42% climb on Qwen3-4B base using our post-training pipeline.

@clo_wang <3

@oroagents GO ORO + TEAM!

All of this work wouldn't have been possible without our partnership with @JarrodBarnes @ Dynamical Systems, and his contributions to the training pipeline. For a detailed breakdown of our training pipeline and the results, refer to our technical report oroagents.com/blog/oro-traje…

For a detailed look at exactly how we are using the data from the subnet in our training pipeline.