Introducing Claude Haiku 4.5

Source: Anthropic Date: 2025-10-15 URL: https://www.anthropic.com/news/claude-haiku-4-5

Summary

Anthropic launched Claude Haiku 4.5, priced at $1/$5 per million input/output tokens. Claims performance comparable to prior-generation Sonnet 4 at one-third the cost and twice the speed. 73.3% on SWE-bench Verified. Surpasses Sonnet 4 on computer use. Positioned as “our safest model yet” by automated alignment metrics. Available via Claude.ai, API, Bedrock, and Vertex AI. Partners: GitHub, Warp, Augment, and Gamma.

Implications

• Claude model cadence / three-tier structure. Haiku 4.5 at $1/$5 establishes the bottom tier of the Claude 4.x family, completing the Haiku/Sonnet/Opus trio at $1/$3/$5. The capability ladder is now explicit: Haiku ≈ prior-gen Sonnet, current Sonnet ≈ prior-gen Opus.
• 73.3% SWE-bench at $1/M. A $1/M input model hitting 73.3% SWE-bench is a significant price/performance milestone — comparable to what premium models cost six months earlier. This compresses the market for mid-tier coding models.
• Computer use leadership. Haiku 4.5 surpassing Sonnet 4 on computer use suggests Anthropic is specifically optimizing smaller models for agentic/automation tasks, not just language quality. Computer use at Haiku speed and cost changes the unit economics for automation workflows.
• Warp and Augment as partners. Terminal-native (Warp) and code review (Augment) partnerships suggest Haiku 4.5 is being positioned for high-throughput, latency-sensitive developer tool integrations where Sonnet is too slow or expensive.
• Watch: whether Haiku 4.5 becomes the default model in Claude Code’s draft/exploratory operations; how open-source models (Llama 3.3, Qwen 2.5) compare at this price point.