Content
summary Summary

Anthropic has introduced Claude Haiku 4.5, a compact AI model that matches the performance of earlier flagship systems while cutting both cost and latency. The company calls it the fastest and most cost-efficient model in the Claude 4 family – and the safest to date.

Ad

Claude Haiku 4.5 is now available worldwide. It replaces Haiku 3.5 and Sonnet 4 as Anthropic’s new standard model and can be accessed through the Claude platform, API, Amazon Bedrock, and Google Cloud’s Vertex AI.

Video: Anthropic

Performance and speed on par with larger models

According to Anthropic, Claude Haiku 4.5 achieves results comparable to the previous top-tier model Claude Sonnet 4, but at one-third of the cost and with more than double the processing speed. In benchmarking tasks like OSWorld’s "computer use" test, Haiku 4.5 significantly outperformed Sonnet 4 (50.7 percent vs. 42.2 percent).

Ad
Ad

In programming benchmarks such as SWE-bench Verified, it reached 73.3 percent – just below Claude Sonnet 4.5 (77.2 percent) but above Sonnet 4’s 72.7 percent.

Image: Anthropic

In tool-use and terminal-coding tests (t2-bench), Haiku 4.5 scored between 41 and 64 percent, a range similar to GPT‑5 and Gemini 2.5 Pro. For math problems in the AIME 2025 benchmark, it achieved 96.3 percent accuracy with Python tools and 80.7 percent without assistance.

Image: Anthropic

The model is tuned for real-time workloads such as chat assistants, customer support, and pair programming. Integrated into tools like Claude for Chrome, it promises lower latency and smoother interactions.

Haiku 4.5 can also run alongside larger Claude series models. For instance, the frontier model Claude Sonnet 4.5 can break down complex tasks into smaller steps and hand them off to multiple Haiku instances for simultaneous execution. This setup allows fast orchestration of task-specific AI agents with lower compute demand.

Pricing is set at $1 per million input tokens and $5 per million output tokens. By comparison, Claude Sonnet 4.5 costs $3 and $15, respectively, while GPT‑5 is priced at $1.25 and $10 per million tokens. GPT‑5 mini remains cheaper, but Haiku 4.5 fills a new middle tier that balances cost and capability.

Recommendation

Secure architecture with low risk classification

Anthropic says Claude Haiku 4.5 underwent extensive safety testing and showed fewer unwanted or misaligned behaviors than previous Claude releases. Statistically, its misbehavior rate is significantly lower than that of both Claude Sonnet 4.5 and Opus 4.1, making it the safest model in the series so far.

Haiku 4.5 launches under the "AI Safety Level 2" (ASL‑2) designation, which carries fewer risk restrictions than ASL‑3 – the category used for high‑capability models like Sonnet 4.5 and Opus 4.1. The company highlights especially low risk in chemical, biological, radiological, and nuclear (CBRN) domains. Details are available in Anthropic’s published Claude Haiku 4.5 system card.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Anthropic has launched Claude Haiku 4.5, a compact AI model that delivers performance comparable to the previous flagship Claude Sonnet 4 while costing one-third as much and running more than twice as fast, now available globally through Claude's platform, API, Amazon Bedrock, and Google Cloud's Vertex AI.
  • The model scored 50.7 percent on OSWorld's computer use test, outperforming Sonnet 4's 42.2 percent, and achieved 73.3 percent on SWE-bench Verified programming tasks, positioning it between Sonnet 4 and Sonnet 4.5, with pricing set at $1 per million input tokens and $5 per million output tokens.
  • Anthropic reports Claude Haiku 4.5 showed fewer unwanted behaviors than previous Claude releases and received an "AI Safety Level 2" classification with particularly low risk ratings in chemical, biological, radiological, and nuclear domains, making it the safest model in the Claude series to date.
Max is the managing editor of THE DECODER, bringing his background in philosophy to explore questions of consciousness and whether machines truly think or just pretend to.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.