← Back to feed

Anthropic halved cache read prices for Claude Sonnet 5.5 on October 7, dropping from $0.20 to $0.10 per million tokens. Standard input and output prices are unchanged. The cut applies immediately across all platforms including AWS, Google Cloud, Azure, and the Claude API.

Why cache reads drive agentic costs

In agentic workflows, the same large system prompt, tool schema, or document context is processed repeatedly across many turns. After the first request caches that content, every subsequent request pays the cache read rate rather than the full input rate. Cache reads make up a large share of total token consumption for long-running agents. At $0.20, they were priced comparably to light input tokens; at $0.10, they become the cheapest component of the Sonnet 5.5 cost structure by a clear margin.

Anthropic estimates the cut reduces typical agentic workflow costs by around 20%. The exact savings depend on cache hit rates: agents with large, stable system prompts that get repeatedly reused see the full reduction; lower-cache or single-turn workloads see proportionally less.

Sonnet 5.5 in the agentic stack

Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 at max effort — the highest score for a Sonnet-class model Anthropic has shipped. Anthropic positioned the cache read cut as a complement to Claude Haiku 5.5 (launched the same day), which targets narrow subagent work. Sonnet 5.5 remains the recommended tier for mid-complexity agentic tasks where Haiku 5.5 is underpowered.

The pricing change makes Sonnet 5.5 more competitive in cache-heavy deployments against alternatives priced at or below its output rate. Existing deployments see the benefit without configuration changes: the API applies the new cache read rate automatically.

Context

Claude Sonnet 5.5 launched on September 28, 2026, making the October 7 cache read cut its first price change since release. The cut applies to Sonnet 5.5 only; other models in the Claude 5.5 family retain their existing cache read pricing.