The Migration
HubSpot migrated Breeze Studio agents to GPT-5 as part of 2026 updates. The headline change: reasoning quality. Multi-step problems, ambiguous inputs, and tool-call planning improved over the prior baseline.
What Got Better
Fewer misinterpretations of user intent. Better tool selection when agents have many actions. More coherent multi-turn conversations — the model stays on task longer without derailment.
What Still Breaks
Hallucinations haven’t disappeared. Policy adherence on restricted topics requires the same prompt-engineering discipline. Cost-per-output is higher on GPT-5 — outcome-based pricing partly hides this, but watch your bill.
Practitioner Action
Re-test agents post-migration. Edge cases that previously failed may now succeed (good) or fail differently (adjust prompts). Don’t assume ‘better model = better agent’ without verification on your use case.