Thanks to @karibriski and Vivek Raghunathan for joining me on stage at Glean:GO last week. Two lines from our conversations have stuck with me. Kari: “Bugs are love letters to my products.” At @nvidia, one of her favorite uses of @Glean is asking which bugs are affecting a product, what’s going wrong, and why. I love the attitude. Don’t hide from the bug queue, it’s one of the clearest signals you have about where the product needs work. Vivek: “There’s a difference between using AI to save yourself 20 minutes and using it to accomplish 80% of your work.” That’s the right bar. The bigger question is what becomes possible when AI takes on most of the work. @Snowflake is working toward that shift, with @Glean helping its engineers find information faster. Thank you both for sharing what you’re learning, and for trusting Glean with part of the work.

We've been investing heavily in our harness and routing capabilities, and we put them to the test benchmarking Glean's token costs against Claude Cowork. The results were striking: @Glean is 4x more cost-effective, averaging $0.45 per task versus $1.84 for Claude Cowork.  That 4x advantage comes from two things compounding: 2.9x lower token volume and a 1.4x cheaper blended rate per million tokens. Here's how:  - Model family routing: Glean made use of Luna which is 10x cheaper than Claude Sonnet and widely capable. We’re able to strike the balance by routing between open and closed models. - Model tier routing: In Glean, Opus was used 10x more (29% vs 2.8%) but surgically for the right things and balanced by other models.  - Better context: Glean’s harness and indexing capabilities result in fewer tokens consumed; Claude Cowork used 3x the tokens per query on average using 88.8M versus 29.8M in Glean.  More results coming out at Glean:GO! Hit me up if you're still looking for an invite.