Episode 78 · Season 1
Jev vs LLMs: Decisions in Milliseconds for 4 Cents per Million Tokens
Every agentic pipeline is full of small yes-or-no and which-bucket decisions, and running each one through a large language model costs seconds and real money at scale. Tim Harting and Gavin Franklin of Cirrius Solutions break down Jev, the new “System One” model from TypeSafe AI that makes those decisions in milliseconds for about four cents per million input tokens.
What you’ll learn:
- Why Jev is a different kind of model, trained with RLCD (reinforcement learning for calibrated decisions), and why it works alongside an LLM rather than replacing it
- The three decision primitives: choose one option from a set, score against up to ten levels, or answer a yes/no question, each with a confidence score you can set thresholds on
- Where it pays off first: content moderation, case and email routing, sentiment triage, model routing, and risk-checking the commands an agent asks to run
Who it’s for:
Engineering leaders, architects, and platform teams who have hit the cost or latency wall when scaling an agentic workflow from one account to thousands.
Guest + credibility:
Tim Harting is COO of Cirrius Solutions and spent the weekend running Jev against his own classification workloads, where several hours of testing landed at a total of about four cents versus the five to eight dollars he would have expected from an LLM. Gavin Franklin is Head of Custom Development at Cirrius Solutions and describes Jev as almost like a logic gate inside the agentic operating system the team has been building for enterprise clients.
Primary CTA:
If you want to talk through where a fast, cheap decision layer fits in your own agent pipeline, reach out to the Cirrius Solutions team.
Recommended Links & Resources:
Related episode: Agentic OS Explained: 4 Pillars That Keep Your AI Agent Sharp
Episode Artifact
TypeSafe AI (Jev)
Tim Harting on LinkedIn
Gavin Franklin on LinkedIn
Greg Banks on LinkedIn
Jason Fowler Music
Cirrius Solutions
Agentic AI Episodes
Any Questions: Reach out to us at CirriusTalk@cirriussolutions.com
We cover
- Exploring the Unique Features of Jeff Model
- Understanding System One vs. System Two Thinking
- JEV’s Decision-Making Capabilities
- Cost and Speed Advantages of JEV
- Practical Use Cases for JEV in Business
- The Role of Jeb in Language Models
Chapters
- 0:00Introduction to Agentic AI and Jeff Model
- 0:28Exploring the Unique Features of Jeff Model
- 0:35Introduction to JEV and Its Significance
- 2:54Understanding System One vs. System Two Thinking
- 5:36JEV’s Decision-Making Capabilities
- 8:10Cost and Speed Advantages of JEV
- 11:13Practical Use Cases for JEV in Business
- 13:36Future Implications and Conclusion
- 24:16Introduction to Cirrius Talk Agency AI
- 24:42The Role of Jeb in Language Models
- 25:04Closing Remarks and Future Engagement
Resources and downloads
The artifacts Greg’s team built for this episode, free to download.
Thinking Fast: Inside Jev, TypeSafe AI’s System One Decision Model
Explains why Jev is a different kind of model, not a smaller LLM: it takes program state and typed questions and returns a typed decision with a confidence score in one fast pass. Covers the economics (about $0.042 per million input tokens and 70 to 500 ms per decision, as reported by TypeSafe AI), the honest limits (narrow by design, text-only, and only as good as the question), and five places a cheap, fast decision helps, including content moderation, LLM output monitoring, model routing, and business process automation.
Related episodes
Agentic AIThe Artifact Ladder: 5 Tiers From PowerPoint to Full AI-Built Apps
Agentic AIClaude vs GPT-5.6: How to Choose the Model Behind Your AI Agents
Agentic AIIs Your Team’s Claude Code Setup the Wild West? Plugins Fix That
The Military MinuteIf Not Me, Then Who? Inside the Travis Manion Foundation
Salesforce at ScalePriscilla Román – From ESL Teacher to Lead Solution EngineerTalk to us about AI agents
Keep going: See how we build AI agents for real businesses, Talk to us about a custom AI agent, More from the Cirrius blog