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.
Indexing is a token cost too. It just does not land in the same column. Glean's 29.8M is measured after a crawl-and-embed pipeline already read the corpus; Cowork has no index, so its retrieval happens inside the query where the meter is running. Part of the 2.9x moved rather than vanished. Churn decides how much. If the embed cost amortizes over enough queries per document, the gap is real and I am wrong to discount it. If the corpus turns over weekly, it is mostly relocation.
How are these tasks characterized in terms of their level of difficulty? Are they delivering the same quality of product / deliverable at the end of task? Gemini models can code too - and for less. I still wouldn't spend any money on them for coding purposes. Other purposes... maybe.
Cost per task is the easy half. The 4x rests entirely on the router spotting which 10% actually needs Opus. Miss that and the tokens you saved come back as a worse answer.
The future of enterprise AI is not just better models. It is intelligent routing that delivers the right model, context, and cost for each task.
Fantastic work, team! You did an incredible job building out the harness and routing logic. The recent integration of GPT-5.6 Luna in Assistant Auto mode on our #Glean instance last week has already been a massive lifesaver for credit usage. Just booked my tickets, see you all at Glean:GO!
4x cheaper from two compounding improvements is a pretty strong argument for treating the harness as part of the product
Guillermo Robatto figured you would find this interesting given our recent discussions.
Love it, Tony Gentilcore!!!! This is a smart way to frame the cost conversation. The 4x number gets attention, but the decomposition is the real story: 2.9x from token volume and 1.4x from blended rate means most of the win is architectural, not pricing!!!! Two things I’d love to see at Glean:GO. First, quality-adjusted cost. $0.45 per task only matters if task completion and accuracy hold at parity, so a cost-per-successful-task view would make this bulletproof. Second, the routing logic itself. Using Opus 10x more but surgically suggests you have a strong signal for when a task actually needs frontier reasoning. That classifier is arguably the most valuable IP here. The context efficiency point is underrated too. 88.8M vs 29.8M tokens says more about indexing and retrieval discipline than model choice. Great work by the team. Can’t wait to catch up at Glean:Go in a few weeks 💙🚀🥳🎉