Sebastian runs Harbor's oracle agent on a Terminal-Bench 2.1 task at 1, 10, and 100 concurrent Tensorlake sandboxes. The oracle agent replays the reference solution, so the numbers measure the infrastructure and not a model. At 100 concurrent, the batch completes in about two minutes: 47 rollouts per minute, a 7.9x gain over 10 concurrent, with 300 of 300 rollouts passed. The usage export puts the cost at 37 cents for each batch of 100, or about $3.70 for 1,000 rollouts.
We didn’t write this one — it’s Sebastian Buzdugan’s piece, published on Data Science Collective. The note above is ours; the full article is theirs.