Anthropic announced Claude Haiku 4.5 on October 15, 2025. The company describes this smaller model as delivering near-frontier coding performance while being substantially faster and more cost‑effective than its recent frontier models. According to Anthropic, Haiku 4.5 provides similar coding capabilities to Claude Sonnet 4 from five months earlier but at roughly one-third of the cost and more than twice the speed.
What changed and why it matters
Claude Haiku 4.5 is positioned for real-time, low-latency use cases such as chat assistants, customer service agents, and pair programming, where responsiveness and cost matter. Anthropic says Haiku 4.5 even outperforms Sonnet 4 on certain tasks like using computers, and it improves responsiveness in applications such as Claude for Chrome and Claude Code.
In their reported benchmarks, Haiku 4.5 achieves about 90% of Sonnet 4.5’s performance on Augment’s agentic coding evaluation while matching much larger models in some settings. Anthropic also reports that the model can run up to 4–5× faster than Sonnet 4.5 at a fraction of the cost, which the company says enables new use cases that were previously impractical due to latency or expense.
Availability and pricing
Claude Haiku 4.5 became available to all users on October 15, 2025. Developers can call the model as claude-haiku-4-5 via the Claude API. The model is also offered on Amazon Bedrock and Google Cloud Vertex AI and can act as a drop-in replacement for Haiku 3.5 and Sonnet 4 at Anthropic’s most economical price point. Anthropic lists pricing as $1/$5 per million input and output tokens.
Safety and classification
Anthropic conducted detailed safety and alignment evaluations on Haiku 4.5. They report low rates of concerning behaviors and state that Haiku 4.5 demonstrated statistically significantly fewer misaligned behaviors in their automated alignment assessment compared with Claude Sonnet 4.5 and Claude Opus 4.1. Because the model poses limited risks for producing chemical, biological, radiological, or nuclear (CBRN) weapons content in their tests, Anthropic assigned Haiku 4.5 an AI Safety Level 2 (ASL-2), whereas Sonnet 4.5 and Opus 4.1 were assigned the more restrictive ASL-3.
The full reasoning and detailed safety test results are published in the Claude Haiku 4.5 system card.
Benchmarks and methodology
Anthropic lists several benchmark results and the corresponding methodologies. Highlights include:
- Augment agentic coding: Haiku 4.5 reaches roughly 90% of Sonnet 4.5’s performance according to Anthropic’s report.
- SWE-bench Verified: Anthropic reports a 73.3% score averaged over 50 trials on the 500-problem dataset, using a 128K thinking budget and default sampling parameters.
- Other evaluations cited include Terminal-Bench, τ2-bench, AIME, OSWorld and MMMLU, with details such as number of runs, thinking budgets (often 128K), and prompt addenda noted for reproducibility.
Anthropic also reports that in one internal instruction-following task for slide text generation, Haiku 4.5 achieved 65% accuracy versus 44% for their premium-tier model, and early tests indicate Haiku 4.5 delivers efficient code generation for GitHub Copilot with quality comparable to Sonnet 4 but at higher speed.
Intended uses and orchestration
Anthropic highlights agent orchestration as a key use case: Sonnet 4.5 can generate high-level multi-step plans and coordinate multiple Haiku 4.5 instances to execute subtasks in parallel. The company positions Haiku 4.5 as a bridge between cost, latency, and high reasoning ability for many development workflows, rapid prototyping, and multi-agent projects.
How to access technical details
Haiku 4.5 is available in Claude Code, Anthropic’s apps, via the API, on Amazon Bedrock and on Google Cloud Vertex AI. Developers and researchers can consult the system card, model page, and documentation for full technical specifications, evaluation details, and safety test results.
