ResolveAI, the San Francisco startup founded by two co-creators of OpenTelemetry, has released a 27B model for diagnosing production alerts and a product that answers the pager on a team’s behalf. The company’s press release, dated November 18, 2026, calls the model SRE-1 and the product Primary On-call, which is generally available now. Founders Spiros Xanthos (CEO) and Mayank Agarwal built it for a job where general-purpose models start blind.

ResolveAI’s argument is about data. Production work has no public training corpus, the company says. Every company runs a different system, incidents stay private, few root causes get documented, and the evidence disappears within weeks. A frontier model arriving at an incident knows nothing about the environment it has been asked to debug.

SRE-1 is ResolveAI’s answer. It is built on open weights and, per the release, was trained by 16 AI researchers across a 7.4-day final run on up to 512 GPUs. Training used reinforcement learning alone. ResolveAI reports that pass rate for one investigation climbed from 6.9% to 38.8% on identical incidents using identical tools. The model is sold only through ResolveAI’s own product.

One design choice deserves attention from anyone who has trained on incident data. Every tool in the training environment was frozen at the instant the alert fired, so the model could not read the outcome from the future. Snowflake’s Vivek Raghunathan, SVP Engineering, says ResolveAI was the first design partner for Snowflake Cortex Training, which was used to build SRE-1.

The headline benchmark is modest in size. ResolveAI took 49 difficult live production cases, ran each five times, and says that pooling five SRE-1 attempts equals the best frontier model it tried, for less money. Xanthos goes further, saying SRE-1 beats every model, frontier ones included, at every budget the company measured. The case set, the grading and the cost figures are ResolveAI’s own. No per-incident dollar numbers appear in the release, and a pool of five runs means the savings depend on how cheap each run is.

A second benchmark covers the whole system. ResolveAI checked 178 alert investigations against verified root causes and says it resolved roughly double the incidents that general-purpose coding agents did on the same alerts. Among the toughest third, its root-cause score ran 15 to 25% higher. The release does not say which coding agents were tested or who verified the root causes.

Primary On-call is where the model meets a pager. It answers the page, resolves the alerts it can, applies mitigations where a team’s policies allow, and hands off to a person when judgment is required. Afterward it compares the result against the production baseline. ResolveAI says the product copes with as many as 80% of pages with no human involved. Customers, according to the company, report ten times faster mean time to resolution and root-cause accuracy of up to 95%. Both are ceilings, not averages.

The most concrete outside number comes from Salesforce. Bhupesh Patel of Salesforce says that during dark-launch testing the product closed over 70% of the company’s pages and no engineer was woken. Zscaler’s Misha Kuperman likewise backs the case for models specialized to a domain. ResolveAI names Coinbase, DoorDash, Figma, MSCI, Nasdaq, Salesforce, SAP and Zscaler as customers.

The rollout is built to be cautious. Teams begin in shadow mode, watching alerts they already understand, and give the product control only of the rules where its record has earned trust. Existing paging-system timeouts stay in place as a backstop, and every decision is logged. That structure lets a skeptical team collect its own evidence before changing anything about who gets woken at 3 a.m.

Shadow mode is the real offer here. Run it against an alert history you can already grade, and the gap between Salesforce’s 70% and ResolveAI’s 80% will show whether the claim holds on your own systems.

ResolveAI press release, “ResolveAI Introduces SRE-1, the Most Accurate Model for Production at Every Budget,” resolve.ai, November 18, 2026.