Every model launch comes with a benchmark chart and a wave of hot takes. What actually matters to an engineering team is much narrower: does it make our product better, and does it do that at a cost we can defend to finance. Now that GPT-5 has been in production long enough for teams to have real data, the picture is more useful than the launch-day headlines.
Where GPT-5 earns its keep
The clearest win teams report is on multi-step coding tasks — refactors that touch several files, migrations that need to track a rename across a codebase, and test-writing that requires understanding existing patterns rather than inventing new ones. GPT-5's longer effective context window means it holds more of a real repository in view before it starts guessing, which shows up as fewer hallucinated imports and fewer broken cross-file references.
Structured output reliability is also better than prior generations. Teams building on top of function calling report fewer malformed JSON responses under load, which matters more than it sounds — a 1% malformed-response rate at scale means thousands of failed requests a day and a retry queue nobody wants to own.
Where it does not obviously win
For short, single-turn tasks — a summary, a classification, a rewrite of one paragraph — teams are not seeing a meaningful quality jump over cheaper models. This is the part that gets lost in launch coverage: frontier models earn their premium on hard, multi-step, high-context work, not on the bulk of everyday requests most products actually send.
Latency is the other honest trade-off. Teams running customer-facing chat report noticeably higher time-to-first-token than lighter models, which pushes some of them toward a tiered routing setup rather than a single model for everything.
GPT-5 vs Claude Sonnet 5: the practical comparison
The two get compared constantly, and the honest answer from teams running both in production is that neither wins outright. GPT-5 tends to edge ahead on tasks that benefit from broad general knowledge and longer planning chains. Sonnet 5 tends to edge ahead on tightly scoped coding tasks and instruction-following where the prompt is precise. The gap on most day-to-day tasks is close enough that existing tooling and API ergonomics often decide the choice more than raw quality does.
Cost per task, not cost per token
The number that actually matters for planning a budget is cost per completed task, not the sticker price per million tokens. A model that costs more per call but needs one attempt instead of three retries can be cheaper in practice. Teams that switched to task-level cost tracking — logging cost against a completed ticket rather than a raw API call — report catching cost blowups weeks earlier than teams still watching aggregate monthly spend.
Practical advice if you are choosing today
- Route by task difficulty, not by habit. Send simple classification and extraction to a cheaper model; reserve GPT-5-class models for multi-file coding, planning, and anything with real business risk attached.
- Test on your own repository, not a public benchmark. The gap between models on your actual codebase, your actual prompt style, and your actual data shape can look nothing like a leaderboard.
- Track cost per completed task, and set an alert threshold before you scale a feature, not after.
- Keep a second model integrated as a fallback. Vendor concentration risk is real, as the Fable 5 export ban showed the industry this summer.
Why this is a staffing question, not just a model question
Picking the right model is only half the job. The harder half is building the routing logic, the eval suite, and the fallback path that makes the choice safe to change later without a rewrite. That is exactly the kind of judgment a dedicated AI engineering team brings to a project instead of a one-off consulting call. If you are building AI features and want engineers who think about model selection as an ongoing decision rather than a one-time pick, see how we staff AI engineering teams or get in touch with your current stack and roadmap.
