Today we’re announcing SRE-1, our first model for finding the root cause of production incidents.
- SRE-1 dominates the cost–accuracy frontier on our benchmark. Across 49 difficult cases from live production, it had the highest pass rate at every budget we measured, starting with its cheapest configuration.
- We built it with a team of 11 researchers. The training run used up to 512 GPUs.
- We started with an open-weight model and used reinforcement learning to train a custom 27B model for production investigations.
More details below.
Introducing SRE-1 a 27B model that dominates the Pareto frontier of accuracy/cost on production debugging, beating frontier models in a coding-adjacent domain
Built by 11 people on open weights models + domain expertise. Final run: 7.4 days on 512 GPUs https://t.co/9IzysHZEmL
How should companies prepare for cyber outages?
"My advice to IT, security leaders and CEOs is that these are very hard domains for building on your own."
"You should find the best companies and models to collaborate with."
"I would say probably you have to collaborate with a different company in coding, a different company in security, reliability, etc. I don't think there's anybody who can do all of that."
"I don't believe the labs by themselves can solve this problem either. They are developing great models, but the last mile complexity in delivering a system reliably and securely is very, very difficult. So you need specialization."
From Industry News with @sophiadew and @resolveai CEO @spirosx
What happens next with cyber outages?
"I think it's only gonna get worse."
"We generate code now at 10x the velocity. We don't have 10x more engineers to run these systems."
"And on top of that, because of all the security vulnerabilities, we are patching these systems now at 10x, 50x the speed than before."
"Unless AI adoption for running production software and solving these problems gets on par with coding and security, we cannot keep up."
"The answer to all of this is AI that can be on the defense, that can help us maintain, patch, and run these systems at the same velocity that we find security vulnerabilities or make changes with AI."
From Industry News with @sophiadew and @resolveai CEO @spirosx
On-call is dead. Let Resolve hold the pager.
• Takes every page, investigates, and mitigates within your guardrails.
• Closes alerts when there’s no impact.
• Escalates to an engineer only when needed.
• Connects alerts to underlying issues and proposes fixes.
Trusted by teams at @Figma, @Salesforce, @Snowflake, and more.
Try Primary On-call today for 3 free months:
https://t.co/bASCmk4rJ2
📟 ♠️ 😇 🔫
22 years after GTA San Andreas, CJ is now a staff software engineer and he's on-call TODAY. Incidentally he's also a Youtuber now, and here's his new vlog on his first ever on-call shift wih @resolveai ! Check it out, like, comment and subscribe!
BREAKING: RESOLVE AI LAUNCHES ITS FIRST MODEL SRE-1
Spiros Xanthos (co-founder and CEO of @resolveai) breaks down the model behind agents that now answer the 3 AM page.
🧠 SRE-1: Opus 5.5-level quality for a fraction of the cost
📟 Primary on-call: Resolve takes the page and escalates only when it needs an engineer.
💻 Why running production is harder than writing code
🌙 What Resolve does in the first two minutes of an alert
🚗 Earning the pager through Shadow Mode and the self-driving analogy
🧑💻 What’s next for SREs and developers
FULL EPISODE BELOW
@spirosx@ravirajjain@lightspeedvp
Today we’re announcing SRE-1, our first model for finding the root cause of production incidents.
- SRE-1 dominates the cost–accuracy frontier on our benchmark. Across 49 difficult cases from live production, it had the highest pass rate at every budget we measured, starting with its cheapest configuration.
- We built it with a team of 11 researchers. The training run used up to 512 GPUs.
- We started with an open-weight model and used reinforcement learning to train a custom 27B model for production investigations.
More details below.
Introducing SRE-1 a 27B model that dominates the Pareto frontier of accuracy/cost on production debugging, beating frontier models in a coding-adjacent domain
Built by 11 people on open weights models + domain expertise. Final run: 7.4 days on 512 GPUs https://t.co/9IzysHZEmL