Short answer: On November 18, 2026, ResolveAI Labs launched SRE-1, a 27-billion-parameter model trained with reinforcement learning (RL) to find the root cause of production incidents. The lab's SRE-1 page says the open-weight model "ties Fable 5.2 at 0.71x the cost." These are the company's own numbers. We found no independent benchmark and no downloadable weights, so treat the headline as a claim to test on your own incidents.
SRE-1 launched with Primary On-call, a product that receives alerts, investigates and acts when policy allows. Below: what the model does, what the company reports, what a relay of the technical paper adds, and the checks to run before you trust it.
TL;DR: what is claimed and what is checked
| Question | Answer | Status |
|---|---|---|
| What is SRE-1? | 27B model trained with asynchronous RL for incident root cause | ResolveAI Labs page |
| How does it compare with frontier models? | "Ties Fable 5.2 at 0.71x the cost" | Company claim |
| How was success scored? | LLM judges score the causal chain at least 0.8 against a reference chain | Company method |
| RL gain? | 6.9% to 38.8% pass on 49 hard live cases | Secondary relay of the paper |
| Are weights public? | Not found on Hugging Face | Unverified |
| Independent test? | None found | Open |
What does SRE-1 do?
The lab's page says SRE-1 "investigates in your environment, working backwards from an alert or issue to find the root cause." It searches through hundreds of tools and integrations and produces a cited causal chain, explaining every step from the root cause to what triggered the alert.
This is a narrower job than a general coding agent. A coding assistant reads code and suggests a fix. An incident investigation depends on telemetry that may expire, queries whose order matters, and actions that can destroy the evidence. ResolveAI argues that this justifies a separate model trained for the job.
Why is production hard for agents?
The page lists four problems, in the company's words:
- A huge search space. Thousands of codebases, petabytes of telemetry and hundreds of tool calls per investigation.
- No easy verification. Most alerts get no postmortem, and some incidents have several root causes or none known, so tests cannot confirm an answer.
- Evidence expires. Telemetry ages out within weeks, so incidents worth evaluating are rare and short-lived.
- Environments leak answers. By evaluation time, the fix is recorded in commits, tickets and chat that a model could read.
These are the reasons a plain benchmark is hard here, and they are the reasons to read the company's numbers with care.
A pager beside a crescent moon representing an on-call engineer and the SRE-1 production incident model
How was it trained?
ResolveAI says it trained SRE-1 with RL on five kinds of environment:
- Live incidents from its own production, labelled for re-investigation.
- Incident replays, with telemetry recorded while an incident is live and replayed so new queries still get answers.
- Synthetic environments built by agent teams, with services, simulated traffic and fake vendors.
- Licensed environments from external vendors.
- Offline synthesis, reaching the root cause from only the saved evidence of an earlier investigation.
A DEV Community write-up of the technical paper, which says it was generated with AI and published without human review, reports a final training run of up to 512 GPUs over 7.4 days. Treat that figure as second-hand until you read the paper.
What are the reported results?
The lab's description says SRE-1 "dominates the Pareto frontier" of pass rate against cost. Pass rate is the share of runs in which LLM judges score the causal chain at least 0.8 against a reference chain, a threshold the company says it calibrated with on-call engineers.
The relayed paper summary adds:
- RL gain. A single SRE-1 investigation passed 6.9% of 49 difficult live production cases before RL and 38.8% after, with the same incidents and tools.
- Ensemble. Five SRE-1 investigations matched the strongest frontier model the company tested, at lower cost. For the cost ratio, the lab's own page gives 0.71x against Fable 5.2.
- Full system. On 178 production-alert investigations graded against verified root causes, the ResolveAI system solved about twice as many incidents as general-purpose coding agents. The reported advantage was 15% to 25% on the hardest third of cases.
All of these are company-reported. A 38.8% single-run pass rate also means most single runs did not pass the bar, which is why the ensemble matters: five runs and a comparison of hypotheses cost more in latency and infrastructure. For a frontier-model view of the same price question, see our Fable 5.2 showcase.
What is Primary On-call and what does it claim?
Primary On-call is the generally available product built around SRE-1. ResolveAI says it handles up to 80% of paged alerts without escalation, and that Salesforce closed more than 70% of paging alerts without waking an engineer during early dark-launch testing, per the DEV write-up of the launch release. When the product cannot safely resolve a problem, it escalates with the investigation attached.
Two reactions appear on the lab's page. Vivek Raghunathan of Snowflake, which supported ResolveAI as the first design partner for Snowflake Cortex Training that it used to build SRE-1, calls it proof of what a domain-specific model can deliver. Zscaler's Misha Kuperman sees "significant opportunity" to work with ResolveAI. Both are partners, not independent testers.
Can you run SRE-1 yourself?
The lab calls SRE-1 an open-weight model. We searched Hugging Face and found no ResolveAI repository, and the model appears to be offered through ResolveAI's product. "Open-weight" may describe the base model it was trained from or a later release. Until a model card and a download link exist, do not plan to self-host it, and ask the vendor in writing what the license and weights policy are.
How does this compare with other small-model claims?
The pattern, a small model trained for one job matching a frontier model at a lower price, has appeared several times this autumn. Coinbase reported a fine-tuned Qwen3.5 9B beating Opus on fraud (our write-up), and Exowatt reported a fine-tuned Qwen3 beating Claude Opus on alarm triage (our write-up). In each case the claim came from the company that built the model, on its own task data. SRE-1 fits the same shape, with a harder task and no public data.
Other AI on-call products are making similar promises; see our coverage of Devin on-call and alert auto-triage, the open-source OpenSRE agent, and a warning about skill atrophy when AI handles incident response.
What does this mean for builders and founders?
For engineers on call. The useful idea is not the parameter count. It is the training setup: replayed incidents, synthetic services and offline synthesis give a model practice on the thing it will face, which is gathering evidence through tools in the right order. If you build your own incident agent, the same recipe applies at small scale: record telemetry during real incidents, replay it, and score the causal chain, not only the final fix.
For founders buying an AI SRE. A vendor with its own model controls cost and data, and can price below a frontier API. The trade-off is lock-in to a model you cannot inspect. The ratio of 0.71x against Fable 5.2 is the type of number that changes with every frontier price cut, so ask for results in dollars per solved incident on your workload, not a ratio against a model that may be repriced next month.
For anyone who writes about it. Say "ResolveAI reports" every time. The company grades its own cases, picks its own threshold, and names its own comparison model. Until an outside team repeats the test, that is the honest frame.
How to test an AI SRE claim before you buy
- Use your incidents. Ask for a trial on replayed incidents from your own stack, with the answers hidden.
- Ask about held-out data. If the vendor trained on its own production incidents, how were test cases kept separate?
- Read the pass definition. A judge score of 0.8 against a reference chain is a judgement, not a fix. Check what an LLM judge accepts.
- Separate alert closing from fixing. A closed page is not a restored service. Track repeat alerts.
- Demand controls. Approval gates, audit trails, rollback and limits on destructive actions are the safety case.
What to watch next
- The weights. A model card, license and download link would let outsiders test the "open-weight" claim.
- Independent results. A third-party benchmark on incident root cause would settle the Fable 5.2 comparison.
- Customer data. Whether more customers publish alert-closure rates like Salesforce's.
How we checked this
We read the ResolveAI Labs SRE-1 page and the paper's landing page, which did not load its full body for us, so paper details come from a DEV Community write-up of the launch. We searched Hugging Face for a ResolveAI model and found none. We did not find the original press release. We will update this post when the weights or independent results appear.
Related reading
- Coinbase's fine-tuned Qwen3.5 9B beats Opus on fraud
- Exowatt: fine-tuned Qwen3 beats Claude Opus on alarm triage
- Devin on-call: what is claimed
- OpenSRE: open-source reliability agent
- AI SRE and skill atrophy
- Small models have arrived
Details are accurate as of November 19, 2026. Figures are company-reported.
