None of the mistakes below are exotic. They are the ordinary consequences of moving fast with a new kind of dependency, and each one is far cheaper to prevent than to unwind. They cluster into three phases — decisions made before writing code, habits formed while building, and gaps discovered in production — so this list is organized the same way.
Planning mistakes (1–4)
1. Committing to a platform before checking feature availability. Teams pick the cloud they already run — a reasonable default — and only later discover the workload needs the Batch API or Files API, which are not available on Bedrock or Vertex AI. The fix costs an hour: list the capabilities your design depends on and check them against the feature matrix before anyone signs or builds.
2. Assuming "Claude 3P" means a different, lesser model. It does not — 3P refers to where you buy and run Claude (Bedrock, Vertex AI, Foundry, Claude Platform on AWS), not a model variant. The same models serve all platforms. Teams that believe otherwise waste evaluation cycles "comparing" model quality across platforms instead of comparing the things that actually differ: features, auth, and commercial terms.
3. Choosing a use case nobody measured. If the current process has no baseline — no time-per-task, no error rate — you will never be able to demonstrate the project worked. Pick a use case with a measurable "before," and measure it before launch. (See measuring ROI.)
4. Bringing security and legal in last. The pilot works, everyone is excited, and then the review that should have shaped the design becomes a blocker at the finish line. A short briefing to security and legal during platform selection is the cheap version; the expensive version is rebuilding on different terms after the fact.
Build mistakes (5–7)
5. Hardcoding model IDs and client setup throughout the codebase. Model IDs differ across platforms (Bedrock prefixes them with anthropic.; everywhere else uses bare IDs like claude-sonnet-5), and every model upgrade or platform move becomes a codebase-wide hunt. Keep the client class, model ID, and region in one configuration point from day one.
6. Defaulting everything to the biggest model. Opus 4.8 is the most capable general model, but a large share of enterprise traffic — classification, routing, extraction, short summaries — runs well on Haiku 4.5 at a fraction of the price, with Sonnet 5 as the balanced middle. Route by task difficulty and let evaluation, not caution, justify the expensive tier. (See model tiering.)
7. Skipping evaluation and shipping on vibes. "We tried it on a few examples and it looked great" is how quality regressions reach customers. A small eval set of real inputs with defined acceptable outputs — run before every prompt change and every model upgrade — is the single highest-leverage engineering habit in this whole list.
Production mistakes (8–10)
8. No cost visibility until the first surprising bill. Token-metered pricing is easy to estimate and easy to ignore. Set budgets and alerts on day one, tag or attribute usage per feature, and watch output tokens specifically — output costs several times more per token than input on every current model.
9. Treating rate limits and transient errors as someone else's problem. Every platform meters throughput, and launch-day traffic finds the ceiling. Implement retries with backoff, sensible timeouts, and graceful degradation before launch, and check your quotas against projected peak load — raising limits takes lead time.
10. Leaving humans out of the loop where mistakes are expensive. Full automation is the right end state for some workflows and a premature one for many. Where an error costs real money or trust — outbound customer communication, compliance-adjacent decisions — keep a review gate until your measured quality earns its removal. The mistake is not automating; it is automating before the numbers say you can.
Where to go next
Most of these mistakes have a full article behind them: the feature gaps for mistake 1, evaluation and testing for mistake 7, and cost alerts and budgets for mistake 8. If you are earlier in the journey, the adoption checklist on the main guide covers the whole sequence.