Opus 5 vs Sonnet 5: A Practical Routing Guide

Published: July 18, 2026 · Updated: October 10, 2026 Read time: 2 min
Opus 5 vs Sonnet 5: A Practical Routing Guide

With Anthropic's 5-series complete — Haiku 4.5, Sonnet 5 at $2/$10, and the new Opus 5 at $5/$25 — the practical question for teams is routing: which model handles which work, and where the premium actually pays.

Opus 5 vs Sonnet 5: A Practical Routing Guide

Update (October 10, 2026): Anthropic's card has dealt again: the current tiers are Opus 5.5 ($4/$20), Sonnet 5.5 (GA September 28, cache reads $0.10 from October 7), and Haiku 5.5 ($0.10/$0.50, October 7). The 5-series pricing below is now last-generation history — the routing logic still transfers.

Anthropic's catalog now spans three clear tiers — Haiku 4.5 at $1/$5, Sonnet 5 at $2/$10, and Opus 5 at $5/$25 per million tokens. The question for engineering teams is not which model is best but which model belongs on which workload. Here is the routing framework we recommend.

Default to Sonnet 5. The balanced tier is the correct starting point for the majority of production traffic: conversational support, code assistance on well-specified changes, summarization, extraction, classification. At half of Opus 5's input price and 40% of its output price, Sonnet 5 is where volume workloads should live unless proven otherwise.

Escalate to Opus 5 on hard steps, not whole pipelines. The flagship earns its premium on tasks with high retry costs: ambiguous architecture decisions, multi-document synthesis, gnarly debugging where a wrong answer costs an hour of human review. The arithmetic to check is retry rate — if Opus 5 solves in one pass what Sonnet solves in three, the premium pays for itself; if both solve in one, you are donating margin.

Reserve Haiku 4.5 for the boring volume. High-frequency, low-stakes tasks — routing, tagging, simple lookups — belong on the cheap tier. Surprisingly often, half of a system's tokens can move down a tier with no user-visible difference.

Cache before you escalate. Before paying flagship prices, check whether prompt caching covers your prefix-heavy traffic. Large stable prefixes — system prompts, tool definitions, document context — can often be served from cache at a fraction of list price, moving the economics more than a tier change would.

Re-evaluate quarterly. Tier boundaries move with every release, and the 5-series is still mid-rollout. The routing table you justify with data this month is worth more than any vendor's marketing claims — keep it current.