Anthropic Unveils Claude Haiku 4.5: Fast, Low-Cost Coding Model

Anthropic launches Claude Haiku 4.5 — faster, cheaper small model for coding and agentic workflows

Anthropic announced Claude Haiku 4.5 on Oct 15, 2025, a new small model aimed at delivering near-frontier coding performance with markedly lower latency and cost. Positioned as a more efficient alternative to its recent frontier releases, Haiku 4.5 targets latency-sensitive applications like chat assistants, customer support agents, and pair programming environments.

Performance highlights and benchmarks

Haiku 4.5 is presented as offering performance comparable to larger models on real-world coding tasks while improving responsiveness. Key claims from the release include:

Coding performance near that of Sonnet 4.5, achieving roughly 90% of Sonnet 4.5’s results in Augment’s agentic coding evaluation.
Substantially faster inference, described as more than twice the speed of Sonnet 4, with cost reductions that make it more economical for high-volume or low-latency scenarios.
Strong results across multiple internal and external benchmarks, including SWE-bench Verified (reported at 73.3% averaged over 50 trials under the specified setup) and various agentic and reasoning suites cited in the model’s methodology.

Anthropic also describes improvements on tasks tied to tool use and agent orchestration, noting Haiku 4.5’s suitability for workflows that split complex problems into parallel subtasks handled by multiple smaller agents.

Use cases and orchestration

Haiku 4.5 is positioned for scenarios where responsiveness matters most. Examples mentioned include:

Real-time chat assistants and customer service agents that benefit from lower latency.
Pair programming and developer-facing experiences where iteration speed is important.
Multi-agent orchestration patterns, where a frontier model like Sonnet 4.5 can produce multi-step plans and Haiku 4.5 instances execute subtasks in parallel.

Safety and alignment

Anthropic reports that Haiku 4.5 showed lower rates of concerning behaviors in automated alignment assessments compared with Haiku 3.5, Sonnet 4.5, and Opus 4.1. Based on the company’s evaluation, Haiku 4.5 has been classified under AI Safety Level 2 (ASL-2). Detailed reasoning and test results are available in the model’s system card: https://www.anthropic.com/claude-haiku-4-5-system-card.

Availability, pricing, and developer access

Claude Haiku 4.5 is available immediately across Anthropic’s apps and developer platforms. Accessibility highlights:

Available via the Claude API (model name: claude-haiku-4-5), Amazon Bedrock, and Google Cloud’s Vertex AI.
Integrated into Claude Code and other Anthropic applications.
Pricing is listed as $1/$5 per million input and output tokens, which is currently midway between OpenAI's gpt-5-mini ($0.25/$2) and gpt-5 ($1.25/$10).

For full technical details, evaluation methodology, and implementation notes, see the model page and documentation: