Glean Says AI Costs 81% Less: The Most Expensive Mistake Is the Denominator

One 81% figure, two entirely different stories

On September 1, The Information distilled the dispute into a sharp headline: Anthropic customers' bills are 80% higher than they need to be. The claim traces to an internal evaluation Glean published on August 26. Across more than 180 enterprise tasks, Glean says its Assistant averaged $0.58 in token cost per task versus $2.98 for Claude Cowork, while reviewers preferred Glean's output 78% of the time.

Calling that 81% cheaper is arithmetically sound: ($2.98-$0.58) ÷ $2.98 equals roughly 80.5%. But with $0.58 as the denominator, Claude costs 5.14 times as much, or about 414% more. This is not pedantry. Every investment percentage begins with a choice of denominator.

The benchmark's most valuable insight is not that it declares a winner. It shows where a new enterprise AI profit pool may emerge. As models become more interchangeable, the control layer that selects a model, compresses context, connects internal data and manages execution may sit closer to the billing gate than the model itself.

From $2.98 to $0.58 measures only one layer

Glean reports that total token use across the evaluated set fell from 4.4 million for Claude Cowork to 1.3 million for Glean, a 70% reduction. Its blended price per token was also 31% lower. Together, those effects produced the claimed 81% reduction in token cost per task. All figures below come from Glean's internal evaluation and have not been independently audited.

The useful reversal: Glean also uses Anthropic

Glean's subprocessor list, updated July 28, names Anthropic alongside OpenAI, Google, Microsoft, AWS, Groq and other model or cloud providers. Glean says its routing evaluates more than 40 open and frontier models, although parts of that capability are marked beta or coming soon; a roadmap should not be counted wholesale as current revenue.

That makes the rivalry more subtle than Glean versus Anthropic. Glean is not proving that Claude lacks value. Its routing may select Claude Opus when the task fits. It is arguing that no single model should monopolize every task's budget. Glean can simultaneously be Anthropic's customer and the gate that controls Anthropic usage.

A durable moat here would come from three feedback loops. More complete enterprise permissions and data relationships reduce wasteful retrieval. More completed tasks improve the router's judgment about when a cheaper model is sufficient. A unified front door can also create purchasing leverage. If customers can cheaply move the same context graph and evaluation history elsewhere, that moat becomes shallow.

The token bill is only the visible tip

Glean explicitly labels $0.58 and $2.98 as token costs. It does not provide a bridge from tokens to a customer's total cost of ownership. Platform or seat fees, indexing and storage, connector maintenance, permission governance, deployment, training, compliance review, human checking, failed-task rework and expected losses from errors can all change the result. Glean also cites its own survey of 6,000 digital workers, saying they spend 6.4 hours per week 'botsitting.' That is company research rather than independent productivity data, but it identifies a cost center that accounting systems often miss.

A better denominator is cost per business-accepted outcome. Full unit cost equals allocated platform and seat fees, plus model tokens, retrieval and storage, integration and governance, human review and rework, and expected error losses. Divide that total only by results that enter the workflow without further repair.

The visible token gap is $2.40 per task. If an employee must spend ten extra minutes checking an output, or a rare error carries a large expected loss, $2.40 can move from headline to footnote. Conversely, if Glean's 78% preference rate translates into materially less rework, labor savings could exceed token savings. Tokens matter; they simply cannot stand in for the entire economic system.

$300 million in ARR proves monetization, not retained value

In a May 28 announcement, Glean said annual recurring revenue reached $300 million, 15 months after crossing $100 million, while its Fortune 500 customer count nearly doubled year over year. That is evidence that enterprises will pay for a context and orchestration layer, and it helps explain why cost routing is becoming a competitive battlefield.

ARR is not profit or free cash flow. The company announcement reviewed here does not disclose audited gross margin, retention, acquisition payback, the portion of token savings passed to customers, or net contribution after model and platform costs. LatePost's August 31 feature, “AI Apps Have Users and Revenue, Yet Are Not Good Businesses”, frames three pressures that apply directly: when the model absorbs the product, whether growth produces profit, and who controls the customer entry point.

If Glean returns most savings to customers, it may increase usage and retention while monetizing less in the near term. If it retains most of the savings as margin, customers gain an incentive to build routing or adopt a neutral gateway. The strongest business shares the gain sustainably: customers can verify net savings, while the platform earns recurring cash tied to accepted outcomes.

Damodaran follows cash flow; Buffett asks whether growth creates value

The next table for investors and CIOs

The most useful future disclosure is not another percentage-savings banner. It is a fixed-cohort table with full-stack cost per thousand accepted outcomes, human rework time, annual net spending and retention 12 months after deployment. Add switching cost, and investors can begin to see whether the context layer compounds.

The bull case is that routing accuracy improves with usage, token and rework costs fall together, and customers entrust more critical workflows to the platform. The bear case is that foundation models bundle enterprise search and routing, connectors standardize, and customers reduce Glean to a replaceable thin layer. The middle case is rapid growth with most efficiency gains transferred to customers to protect the entry point.

Use Valuefocus's reverse DCF to test what revenue and cash flow the price requires, then record model substitution, switching cost and vendor-evaluation bias in an investment risk checklist. The most expensive part of enterprise AI may not be the price per million tokens. It may be mistaking the easiest denominator to display for the economics of the whole business.