DevRev delivers autonomous self-service with GPT-6.1 Sol

OpenAI published a customer story on DevRev, a North America tech startup using GPT-6 Astra and GPT-6.1 Sol to power Computer and Operator for customer and employee service. The pitch is autonomous self-service that still respects enterprise constraints: reason over live data, keep actions reversible, cite sources, and enforce what an agent is allowed to do.

CTO Ahmed Bashir's bar is blunt — right reasoning, trusted and aligned behavior, and latency that works for real users. That triad is the enterprise adoption story in miniature. Models that only win on benchmarks do not ship into support queues.

This is a vendor case study, so we treat performance claims as DevRev and OpenAI's account, not an independent audit.

We like where DevRev is aiming even while we discount the glossy framing. Enterprise agents fail in the gap between a smart reply and a permitted action on messy, multi-system data. Bashir describes testing models together with the harness that supplies company context, enforces permissions, and controls tools. Their Enterprise-Bench reportedly stresses versioned data, information split across systems, and role-based access — the ugly conditions production actually has.

GPT-6.1 Sol, in DevRev's telling, hits the balance of nuance, speed, and sustainable cost when a customer or employee is mid-conversation. "I think it strikes the balance perfectly," Bashir says. We're glad the optimization target includes cost and latency, not only peak reasoning. Support workflows die when every turn needs a deep think.

The Codex detail is a quiet industry tell. Bashir says the whole engineering team now uses Codex and dropped Cursor licenses after trying multiple coding tools — "Developers are really just looking for the best tool, and you can't fool them." Sales engineers also use GPT-6 Astra for conference demos. That is one company's preference in a fast-moving tooling market, not a verdict on every IDE, but it shows how quickly teams will rip out seats when something feels better.

I'm watching the next ambition more than the launch copy. DevRev wants Operator to take longer, harder requests while staying traceable, permissioned, and reversible — beyond knowledge-base answers, plan upgrades, and refunds into true self-service. Bashir wants more of that logic described conversationally, with the right intelligence for each step, completing "in seconds, not minutes." Useful if they hold the line on reversibility and permissions. Risky if "autonomous" outruns audit trails.

Advocate for the direction: agents that finish work inside company policy are how AI becomes operationally useful. Name the downside: customer stories do not show failure rates, escalation quality, or what happens when the model is confidently wrong with write access. Those are the metrics that matter once Operator touches refunds and account changes at scale.

DevRev combines OpenAI models with its own harness and Agent Studio, where users can describe tool behavior in conversation and lock parts down with fixed rules. OpenAI lists products used as API and Codex. Bashir also eyes OpenAI's Decisions API and Ultrafast options for steps that should not require deep reasoning every time.