DevRev builds autonomous enterprise self-service on GPT-6.1 Sol

Hi, it's Shiichan! I read an OpenAI customer story today and want to share it with you.

OpenAI News openai.com

What was announced?

OpenAI's News has a story about DevRev using GPT-6.1 Sol to deliver autonomous self-service for enterprises.

DevRev built two products, Computer and Operator, to help service and operations teams complete difficult customer and employee requests.

Why it matters

This is a customer story rather than a new feature announcement, but it's interesting because it shows what an enterprise weighs when choosing a model for AI agents.

What changes

According to the article, the two products split the work like this:

• Operator: understands an issue and takes approved actions
• Computer: receives requests that need a person from Operator and prepares the right person to act

Both are designed to reason over live data, actions are reversible, and answers link back to their sources.

Dive Deep

On model choice, the article says:

• OpenAI models, including GPT-6 Astra and GPT-6.1 Sol, power the reasoning, planning and tool selection in Computer and Operator
• DevRev's CTO, Ahmed Bashir, says the models offer reasoning and reliability at a speed and cost that suit customer and employee service
• Evaluation uses DevRev's own Enterprise-Bench, which tests data with multiple versions, information spread across systems, and role-based access

On the architecture side, DevRev combines OpenAI models with its own harness. The harness connects company knowledge and controls what actions an agent can take. In Agent Studio, users can describe a tool's behavior conversationally and set fixed rules for parts of it.

Other uses are mentioned too:

• Codex: the entire engineering team uses it, and the company moved fully to Codex and dropped its Cursor licenses
• GPT-6 Astra: sales engineers use it to build demos for conference booths
• Next: DevRev sees potential for the Decisions API and Ultrafast to speed up workflows that don't need deep reasoning at every step, and wants Operator to handle longer, more complex requests while keeping actions traceable, permissioned and reversible

The article gives no quantitative results such as metrics, benchmark scores, cost or latency. Tasks like knowledge-base lookups, plan upgrades and refunds are described as a starting point, not achieved milestones.

Wrap-up

• DevRev uses Computer and Operator to handle hard requests for service and operations teams
• GPT-6 Astra and GPT-6.1 Sol power reasoning, planning and tool selection
• The deciding factor is the balance of reasoning, speed and cost, evaluated with Enterprise-Bench
• The article reports no numeric results

This is for anyone building enterprise agents who wants to see how model choice and safe action design are approached!