Opus 5.5 Pricing Explained: The 20% Cut Is Only Half the Story

Claude Opus 5.5's pricing is more interesting than the headline 20% cut suggests: $4/$20 per million tokens plus cache reads at $0.20 — a combination that reprices the repetitive-prefix workloads that dominate production traffic.

Opus 5.5 Pricing Explained: The 20% Cut Is Only Half the Story

The headline on Claude Opus 5.5's pricing is a cut of roughly 20%: $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25. But the changes that will actually move most production bills are in the caching line items.

Prompt cache reads on Opus 5.5 cost $0.20 per million tokens — 60% cheaper than on Opus 5, Anthropic says. Cache writes cost $5 per million tokens for five-minute retention or $8 for one-hour retention. For workloads that reuse large, stable prefixes — system prompts, tool definitions, documentation corpora, codebase context — this is where the money is.

Consider the shape of a typical production workload: a coding assistant that sends a 100,000-token repository context with every request. Under list pricing, that context dominates cost. With a 90% cache hit rate, the effective input price on that prefix falls from $4 per million to $0.20 per million — a 20x reduction that dwarfs the headline cut.

The rest of the Claude lineup now brackets Opus 5.5 cleanly: Sonnet 5.5 at $2/$10 for balanced work (cache reads down to $0.10 from October 7), Haiku 5.5 at $0.10/$0.50 for high-volume simple tasks after Haiku 4.5's October retirement. Against Google's promotional Gemini 3.8 Flash at $0.75/$3.75, Anthropic is not the budget option — the pitch is capability and caching economics for serious workloads rather than raw floor pricing.

The broader trajectory matters more than any single price. Two frontier labs cut prices this month while a third introduced promotional rates, all under pressure from MIT- and Apache-licensed open-weight alternatives. Model costs are falling on a trend line, and the engineering practice of routing workloads between cache-friendly premium models and cheap tiers has become a core competency.