Claude Haiku 5.5 ships ~75% cheaper than Haiku 4.5; Sonnet 5.5 cache reads halved

Anthropic completed its Claude 5.5 family with Haiku 5.5, its fastest and cheapest model. Reuters notes it is the third 5.5 release in a month and arrives ahead of Anthropic's planned IPO. The company says Haiku 5.5 runs about 75% cheaper than Haiku 4.5 on most tasks. Target workloads include classification, summarization, extraction, live support, voice, and subagent work under Opus 5.5 or Sonnet 5.5. It launched on Anthropic's platform and on Amazon Bedrock the same day.

The second lever is caching. Sonnet 5.5 cache reads drop 50%, from $0.20 to $0.10 per million tokens. For agent loops that resend large system prompts and tool definitions every turn, cached input dominates the bill, and Anthropic estimates overall agentic costs fall about 20%. Bedrock simultaneously expanded explicit prompt caching. Earlier in the cycle, Sonnet 5.5 came in 30% cheaper than Sonnet 5 and Opus 5.5 40% cheaper than Opus 5.

The competitive frame is an open pricing war. Yahoo Finance headlined the 'AI pricing war'. Grok 4.5 sits at $2/$6 per million tokens, Microsoft is moving Copilot to pay-per-use, and DeepSeek's V4.1-Flash claims a lead in agentic coding. On HN, Haiku 5.5 hit 679 points, and developers are re-running cost-per-task spreadsheets.

Caveats: '75% cheaper' depends on workload mix, and small-model quality on long multi-step tasks still needs independent benchmarks. Separately, the community is watching reports that Meta and Microsoft are reducing internal Claude usage on cost grounds. Aggressive pricing may be partly defensive.