The setup
"Agent" here means: a model that calls tools (functions, APIs, code interpreters), reads the results, decides what to do next, and either continues or returns. We're not talking about chat with retrieval bolted on; we're talking about loops that can take 10-50 model turns to complete a single user task.
Across our client work, we've shipped agents in this shape on both Claude and ChatGPT. Both models are good enough to be the brain. Failure modes diverge under load, though, and the divergence tells you a lot about which one to pick.
Where Claude breaks first
Over-caution refusals on benign tasks. Claude will occasionally refuse or hedge on a tool call that touches anything it pattern-matches as sensitive — even if the user context makes the task obviously fine. The frequency has dropped a lot since Sonnet 4.x, but not to zero. Our workaround: explicit user-context framing in the system prompt and a one-shot example showing the model what the right behaviour is. This solves 90%+ of cases.
Rate limits at peak. Anthropic's frontier-tier capacity tightens during US business hours. Bursty agents hit it. Mitigation: prompt cache aggressively (drops effective rate by 5-10x on input-heavy turns), and have a fallback provider — we route to Bedrock or Vertex when the direct API throttles.
Long agentic loops can stall on a sub-step. Claude sometimes asks the user a clarifying question instead of taking a reasonable default. This is usually the right behaviour but not always — for fully autonomous agents, you sometimes need to nudge the system prompt to "if uncertain, pick the most reasonable interpretation and continue."
Where ChatGPT breaks first
Schema drift. GPT models are less disciplined about sticking to a declared JSON schema than Claude. We've seen extra keys appear, types change, optional fields go missing. Our workaround: strict schema validation on every output, with a single retry loop on validation failure. This is fine but it's overhead Claude rarely needs.
Model deprecations. OpenAI ships fast, deprecates fast. We've had to migrate production code three times in 18 months. Each migration was minor in isolation; the cumulative engineering tax is real. Anthropic's deprecation cadence is slower and more predictable.
Tool-call hallucination. GPT-5.5 occasionally invents tool parameters or calls tools that don't exist in the provided schema. Our agents handle this by silently dropping invalid calls and re-prompting, but it adds latency and shows up in our error rates as a specific class of bug we don't see on Claude.
What this means for picking
For agents that have to be reliable in tool use — anything calling real APIs, anything writing to a database, anything in customer production — we default to Claude. The fewer schema errors and cleaner refusal-instead-of-confabulation behaviour pays for itself in fewer 3am pages.
For agents where ecosystem breadth matters more than tool reliability — a research assistant that should access many third-party services, a consumer-facing agent that benefits from OpenAI's plugin ecosystem — ChatGPT is often the right call.
A specific note on cost: don't pick on per-token price alone. Claude Sonnet at $3/$15 is more expensive than GPT-5.5 at $5/$30 only at the surface — once you factor in fewer retries, fewer schema-validation loops, and fewer human-in-the-loop escalations from confused tool calls, Claude usually wins on net for tool-heavy workloads.
What we keep doing on both
- Strict schema validation on every tool-call output, regardless of model.
- Cap on loop iterations (15-25 turns) with explicit human escalation.
- Cost dashboard per agent run — outliers flag for human review same day.
- Eval set replay weekly to catch regressions when we change prompts or upgrade models.
Cross-references
- Full Claude review with model tier breakdown and pricing math.
- Full ChatGPT review.
- Claude vs ChatGPT — head-to-head spec comparison.
- LLM cost calculator — model your actual workload.