Hi, this is Ravit.

Every enterprise AI conversation I have had this year eventually lands on the same question. Not which model is smartest. Not which agent demo looks best. The question is what this actually costs when you run it at scale, every day, across thousands of tasks.

This week Glean put a number on it. The team benchmarked its own agentic platform against Claude Cowork, Anthropic's agentic knowledge work app, and published the results. The claim is direct: Glean averaged $0.45 per task. Claude Cowork averaged $1.84. That is a 4x difference on the exact same category of work.

I went through the numbers Glean shared, including the breakdown from the co-founder's post, and did some math of my own on top. Here is the full picture, what is driving the gap, and the questions I would bring to my own evaluation.

THE HEADLINE NUMBERS

Two charts tell most of the story. Cost per task and tokens per task.

Per Glean's benchmark run, Claude Cowork consumed roughly 3x the tokens per query on average. Across the full test, that was 88.8M tokens for Cowork versus 29.8M for Glean. Same work, very different meter readings.

Token volume is only half of the equation. The other half is what each token costs you. And here is a calculation worth doing yourself, because Glean did not spell it out. Divide cost per task by tokens per task. Glean lands at roughly $0.65 per million tokens blended. Cowork lands at roughly $0.92. That is the 1.4x rate advantage.

Multiply the two factors together, 2.9x fewer tokens and a 1.4x cheaper blended rate, and you arrive at the 4x headline. The gap is compounding, not additive.

WHAT IS DRIVING IT

Glean attributes the gap to three engineering decisions. Each one maps directly to how any enterprise should think about agent economics.

Glean does not run every task through one frontier model. It routes across open and closed models based on what the task actually needs. A key piece here is Luna, which Glean says is 10x cheaper than Claude Sonnet while remaining widely capable. Cheap, capable models handle the volume. Frontier models handle the moments that justify frontier prices.

This is the counterintuitive part. Glean actually used Opus, the most expensive tier, far more than Claude Cowork did. 29% of Glean's token volume versus 2.8% in Cowork. The cheaper system used 10x more of the priciest model.

That only makes sense if you optimize cost at the task level, not the token level. Glean deployed Opus surgically, on the steps where top-tier reasoning changes the outcome, and balanced it with cheaper models everywhere else. Spending more on the right tokens made the whole task cheaper.

Glean's harness and enterprise indexing feed the model tighter, more relevant context. This is retrieval precision showing up on the meter. Prompts carry less dead weight, trajectories converge in fewer turns, and the token count drops. This is where a decade of search and knowledge graph work quietly pays off in the agent era.

WHERE THE TOKENS WENT

The model distribution chart is the most revealing part of the benchmark. It shows two completely different philosophies.

Claude Cowork ran 96% of its tokens on the Sonnet family. Essentially a single point on the capability-cost curve. Glean spread its volume across the GPT-5.6 family at 64%, Opus at 29%, and other models at 7%. A portfolio spread across the frontier.

The takeaway is not that any single model is bad. It is that a single-model strategy leaves optimization on the table in both directions. You overpay for easy steps and underpower hard ones.

Here are the original charts Glean shared with the benchmark:

WHAT THIS MEANS IF YOU ARE BUYING

At small scale, the difference between $0.45 and $1.84 is noise. At enterprise scale it is a budget line. Run 100,000 agent tasks a month and the gap is roughly $139,000 per month, or about $1.7M a year, on Glean's numbers. That is the kind of delta that decides platform strategy, not just tooling preference.

The bigger shift is what gets benchmarked. For two years the industry compared models on intelligence. Now the comparison is moving up a layer, to the harness. Routing, context engineering, and indexing are becoming the differentiators, because they determine what the same intelligence costs to deploy.

HOW I WOULD READ THIS BENCHMARK

As always, this is a vendor-run benchmark, so bring the same questions you would bring to any of them. Ask what the task mix was and how closely it matches your workloads. Ask how quality was scored alongside cost, because a cheap wrong answer is the most expensive output there is. And run your own pilot on your own data.

Glean says more results are coming at Glean:GO, its upcoming event, so the methodology conversation is about to get deeper. That is exactly the direction this space needs: vendors publishing numbers and being ready to defend them.

See how Glean's Work AI platform handles routing, context, and cost at enterprise scale.

The Ravit Show Data and AI, explained through the people building it. 1M+ community | 137K+ newsletter subscribers | 750+ interviews

This edition is in partnership with Glean. Benchmark figures are from Glean's internal testing.