Faros

We tested whether GLM 5.2 could replace Opus 4.8 for a meaningful share of Claude Code traffic. Before the pilot started, we set a clear bar for success: the route had to deliver significant cost savings without creating enough workflow friction to undermine them. Five days of real developer usage gave us a pretty clear answer, but it also surfaced an issue we wouldn’t have found by looking at cost or model quality alone. We break down what happened, what we learned, and the rollout decision we made here: https://lnkd.in/g_E_cvD9

Chase Norton, our Head of AI, is joining the AI Panel at TestMu Conf 2026 alongside speakers from Microsoft, Qodo, and Moderne. They'll get into validating AI-generated changes, scaling standards across repositories, and deciding where deeper automation is safe to trust. Registration is free, sign up today:  https://lnkd.in/gKubcJU9 #TestMuConf2026 #AIEngineering

Your engineering team can hit every AI efficiency signal in the green—right model, solid plan, guardrails in place, clear definition of done—and still ship nothing that moves the business forward. With AI coding bills skyrocketing, leadership needs to know what's resulting from AI initiatives, but most teams can't explain much beyond a token spend chart. Join us tomorrow as Chase Norton (Head of AI) and Saba Mahdavi (Forward Deployed Engineer) explain how to close that gap. You'll leave with a framework for attributing AI coding spend to real outcomes, even without a clean data pipeline. 📅 Wednesday, August 12 🕙 10:00–10:45 am PT Save your seat: https://lnkd.in/gcgUaFd4

An AI coding session can check every efficiency box and still deliver nothing that matters. Chase Norton and Saba Mahdavi break down how to trace AI coding spend to real outcomes, even without a perfect data pipeline. Join us on Aug 12 from 10:00–10:45am PT. Register here 👉 https://lnkd.in/gcgUaFd4

If you're at Ai4 in Las Vegas this week, don't miss our talk, "Adoption Is Solved. AI Engineering Is Not." this Thursday on the AI ROI track at 11am PT. Drop by booth #1461 and chat with our experts about what it actually takes to turn AI coding spend into measurable outcomes. 📸 feat. Chase Norton, Evan Bruns, and George Youhana

Join us in one hour for our Tech Talk on Token Intelligence to learn how engineering leaders are gaining visibility into productive, inefficient, and wasteful AI coding spend. Register: https://lnkd.in/gJUy9pcT

Is intelligent model routing enough to improve AI coding performance? Our latest experiment suggests routing decisions need much more context. We ran 211 real engineering tasks through six model-and-harness combinations. The aggregate-best route led on average quality, cost, and speed, yet it was the best choice for only 40% of tasks. Other routes performed better for the remaining 60%. The best choice changed by repository, work type, and complexity. On average, the quality gap between the best and worst route for a task was 43 points. Read the full analysis: https://lnkd.in/e8BcJ8c4

Your AI session was efficient, but did it deliver any value? We're gearing up for the second webinar in our latest series: From Token Spend to Outcomes Join Chase Norton and Saba Mahdavi on Aug 12 to learn how to connect AI coding spend to results your leadership actually cares about. Register: https://lnkd.in/gcgUaFd4

What would it cost to scale your best engineers' AI usage across the org? Next week, Evan Bruns, Forward Deployed Engineer at Faros, will show how engineering teams are connecting AI spend to outcomes and identifying where AI is creating value. Register here: https://lnkd.in/gJUy9pcT

In case you missed it, our recent webinar "What is Productive AI?" is now available on demand! A highlight: Chase Norton tackled the same bug two ways. The efficient approach was nearly 5x cheaper and scored 35 points higher on PR quality. If your team is spending more on AI tokens without better outcomes, this one's worth the watch! Watch now: https://lnkd.in/gDhr6yqi