Choosing the Claude model "just in case, go with the most expensive one" is as common a reasoning error as always picking the cheapest — both ignore that real cost isn't just price per token.
Haiku 4.5 is priced at $1/$5 per million input/output tokens — fast and cheap. Sonnet 5 at $2/$10, the balanced option. Opus 5 at $5/$25, the recommended choice for complex agentic work. Fable 5, the most capable model Anthropic offers, is priced higher still (see the dedicated article comparing that specific case).
Haiku 4.5: route routine classification and high-volume work to Haiku — it's the right choice when the task is simple and repetitive and volume makes every cent per call matter. Sonnet 5: the default recommendation for most production use cases — it delivers quality close to Opus with faster latency and significantly lower cost. Opus 5: at $5/$25, it's the sweet spot for most production work requiring complex agentic reasoning, multi-step code, or tasks where a wrong answer is expensive.
Opus 5's efficiency gains — fewer retries, fewer tokens, better self-verification — mean it can end up cheaper in practice than a lower-priced model that needs more iterations to reach the same correct result. Comparing only the price per million tokens without considering how many total tokens (including retries) each model needs to complete the task is a common estimation mistake.
Cache hits cost 10% of the standard input price — for any architecture with repeated context across calls, this is the highest-impact optimization available. The Batch API cuts input and output prices by 50% — for work that doesn't need a real-time response, that's directly half the cost, regardless of which of the three models you chose.
Start designing your architecture with Sonnet 5 as the sensible default, measure real quality against your own test cases, and only move up to Opus 5 for the specific sub-flows where Sonnet falls short or where the cost of an error is high. Reserve Haiku 4.5 for the classification or simple-extraction sub-flows you identify within the same pipeline — the optimal architecture almost never uses a single model for everything, it mixes all three according to the real demands of each step.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel