[DigitalToday reporter Chi-gyu Hwang] AI search startup Glean is mounting an offensive against Anthropic, taking aim at token costs and data security issues.

On Sept. 1 local time, The Information reported that Glean last week released the results of an internal evaluation showing its AI assistant was much cheaper than Anthropic’s Claude CoWork app when carrying out more than 180 tasks.

Glean provides tools that help customers such as Intuit, Booking.com and Zillow build data context layers. This includes an enterprise graph that helps AI understand connections across corporate data.

A Glean official said the company’s assistant, powered by Anthropic’s Claude Opus 4.8 and other models, used 70 percent fewer tokens when performing the same tasks. The comparison was against Claude CoWork powered by Claude Sonnet 5 in high reasoning mode.

Glean also stressed that its average charge per task was 81 percent lower than Claude CoWork. The company said this was because it also used its own routing software that allocates tasks to different models depending on complexity. Anthropic did not issue a separate response.

Glean CEO Arvind Jain (아빈드 자인) said, "The context layer reduces the cost of the search process for AI to find the right data." He said, "Until last year, AI did not do complex tasks so costs were not that big, but now it performs complex tasks, so it takes a long time and costs a lot."

As of May, Glean’s annual recurring revenue had exceeded $300 million. Glean is competing in the enterprise AI agent market with major cloud providers as well as Anthropic and OpenAI.

Glean stressed that its agents can operate even under the security standards and data protection environments that companies apply to employees. A Slack-based agent that Glean launched in June replaces engineers who stay on constant standby to keep applications running. If an application goes down, the agent gathers information needed to identify the cause and suggests solutions.

Jain said Glean is also working to give each agent a separate identity and authentication information so it can access only authorized data. He said, "Agents that run for a long time and have a lot of privileges and the ability to perform tasks in systems can easily slip out of control."