We deployed Claude Sonnet 5 into our AI agents. Here's what actually changed.

Our team spent the past week rolling Claude Sonnet 5 into a few AI agents we've been building for enterprise clients. Went in a bit skeptical honestly, most model updates lately have felt like marginal wins that don't really move the needle in production. Sonnet 5 surprised us in a good way.

What we tested it against

• Multi-step research agent: Pulls data from 6+ sources, summarizes, cross-references, and generates a report

• Coding agent: Handles refactoring tasks and PR reviews across a mid-sized codebase

• Customer support triage agent: Classifies tickets, routes them, and drafts first-response messages

What actually changed:

• Planning quality is noticeably sharper. Sonnet 5 breaks down tasks into cleaner sub-steps than Sonnet 4.6 did. Our research agent used to occasionally skip a source or misinterpret the ordering of steps, now it's producing consistent, well-structured plans on the first pass. Way less babysitting from us during runs.

• Tool use feels more reliable. Fewer weird tool-calling errors, better recovery when a tool returns an unexpected format. This alone saved us hours of debugging over the week. The coding agent especially benefited, PR review comments went from "generic and obvious" to "actually useful, sometimes catching subtle issues we missed."

• Latency and token usage: Slightly more verbose than 4.6, but the intro pricing at $2/$10 per M tokens is honestly a steal. Even at standard pricing after promo ($3/$15), we're likely going to keep it as the default for most workflows.

• Compared to Opus 4.8: For the tasks we tested, Sonnet 5 delivered maybe 90% of the quality at a fraction of the cost. Opus still wins on the hardest reasoning tasks, but for daily agentic work, Sonnet 5 hits the sweet spot.

The question I want to throw out:

For anyone else running AI agents in production or in build stage:

• Have you tested Sonnet 5 in your workflows yet? What did you notice?

• If you've also tried the latest GPT model updates, how does Sonnet 5 stack up in real agentic tasks?

• Any use case where you found Sonnet 5 actually underperformed compared to what you had before?

Would love to hear real production experiences, not just benchmark takes. Trying to decide whether to fully migrate our stack or keep a hybrid setup across models for different task types.