DevRev bets on GPT-6.1 Sol for autonomous customer service

DevRev is running its autonomous customer-service agents, Computer and Operator, on OpenAI's GPT-6.1 Sol, a model it says balances reasoning quality, response speed, and cost for enterprise service workloads.

The company positions the pairing as the foundation for what it calls autonomous self-service: AI agents that carry customer and employee requests through to completion rather than handing them off to humans. Whether that promise holds at scale remains the open question — but DevRev's internal engineering choices suggest it is betting heavily on the OpenAI stack.

What are Computer and Operator?

Computer and Operator are DevRev's two agent products, built to help service and operations teams complete even the hardest customer and employee requests. According to Ahmed Bashir, DevRev's chief technology officer, the agents are designed to reason across live company data. He stressed two control mechanisms: "Actions are reversible, and answers point back to their sources."

The two agents divide labor. When a customer or employee asks for help, Operator may need to understand the issue, take an approved action, or route the request to Computer, which prepares the right person to act. The architecture presumes that not every step requires deep reasoning — a design choice aimed at keeping latency and cost in check.

How does DevRev evaluate the models?

DevRev does not rely on public benchmarks alone. It tests OpenAI models — including both GPT-6 Astra and GPT-6.1 Sol — together with the surrounding software that supplies company context, enforces permissions, and controls agent actions.

Its internal benchmark, called Enterprise-Bench, evaluates models under conditions that mirror real enterprise environments:

• Data exists in multiple versions simultaneously
• Relevant information sits across different systems
• Access to that information depends on a person's role

Ahmed says GPT-6.1 Sol performs well in those evaluations. The reasoning is practical: when a customer or employee interacts with an AI agent, the model must grasp the nuances of the request, respond quickly, and do so at a cost the business can sustain. "I think it strikes the balance perfectly," Ahmed says.

That is a vendor's assessment of its own deployment, and no independent benchmark results were published. What DevRev does offer is a named evaluation methodology tied to concrete enterprise failure modes.

Why did DevRev drop Cursor for Codex?

Inside DevRev, engineers have made a decisive tooling choice. "Our entire team uses Codex," Ahmed says. The adoption grew organically after engineers explored multiple coding tools. By the time of the interview, DevRev had moved fully to OpenAI's Codex and dropped its Cursor licenses.

For Ahmed, the shift reflects what developers actually value in a tool. "Developers are really just looking for the best tool, and you can't fool them," he says. His engineers were unequivocal: "We think that Codex is far and away the best experience."

The detail matters for the broader AI coding-tools market, where Cursor and Codex compete for engineering teams. A full migration by an AI-native company — one with the expertise to evaluate the options — is a stronger signal than a procurement decision made on price alone.

DevRev's sales engineers also use GPT-6 Astra to build demos for the company's booths at international conferences.

What comes next: Decisions API and Ultrafast

Ahmed sees opportunities for OpenAI's Decisions API and Ultrafast to help DevRev's workflows finish quickly, without requiring deep reasoning at every step. The reasoning is economic as much as technical: some decisions are difficult and expensive to define with fixed rules, especially when the information those decisions depend on keeps changing.

Ahmed wants builders to describe more of that logic conversationally and pick the right kind of intelligence for each task — heavy reasoning where it earns its cost, fast lightweight inference everywhere else. "These things have to be able to run and complete themselves in seconds, not minutes," he says.

In DevRev's Agent Studio, that conversational approach is already partially in place. Users can describe how they want a tool to behave in a conversation with AI, then pin parts of that behavior down with fixed rules — a hybrid of natural-language specification and deterministic guardrails.

Can self-service go beyond refunds?

DevRev's stated next ambition is for Operator to take on longer, more demanding customer and employee requests while keeping actions traceable, permissioned, and reversible. Those three constraints — traceability, permissions, reversibility — are the enterprise governance layer that separates agentic AI from a demo.

Ahmed frames familiar support tasks such as finding a knowledge-base answer, upgrading a plan, or getting a refund as only the beginning. "We want to really realize that promise of self-service," he says.

The stakes for the enterprise software market are straightforward. Customer service is one of the largest and most measurable cost centers in any company, and autonomous resolution — not deflection to help articles, but completed actions — is the metric that would justify AI spending at scale.

What to watch

DevRev is combining OpenAI's models with its own harness for connecting company knowledge and giving teams control over what agents may do. OpenAI models power the reasoning, planning, and tool selection behind both Computer and Operator; DevRev supplies the context, permissions, and action layer.

"It's a combination of being able to have the right reasoning capabilities, being able to be trusted, highly aligned, and most importantly, it has to work from a latency perspective," Ahmed says of the model choice.

The claim that agents can complete complex, multi-step requests autonomously — in seconds, with reversible actions — is the promise. The Enterprise-Bench results and the full migration to Codex are the evidence DevRev has offered so far. The next test is whether Operator can sustain longer, harder requests at production scale without eroding the traceability and permissioning the company has built in.