Anthropic has released Claude Haiku 4.5, a latency-optimized small AI model. It matches the coding performance of Claude Sonnet 4 at more than double the speed and one-third the cost, and is available via API and cloud partners.
The model targets applications with tight latency budgets and high throughput demands, including real-time assistants, customer-support automation, and pair-programming. Anthropic positions Haiku 4.5 as a drop-in replacement for Haiku 3.5 and Sonnet 4 in cost-sensitive, interactive workloads. It reportedly surpasses Sonnet 4 on “computer use” tasks, which involve GUI and browser manipulation for products like Claude for Chrome. The model is also described as materially improving responsiveness in Claude Code for multi-agent projects and rapid prototyping.
In Anthropic’s model hierarchy, Haiku 4.5 provides near-frontier capabilities with greater cost-efficiency. The company affirmed that its Sonnet 4.5 model remains the frontier offering and “the best coding model in the world.” Anthropic suggests an orchestration pattern where Sonnet 4.5 handles complex, multi-step planning, and a pool of Haiku 4.5 workers executes the sub-tasks in parallel. This approach is recommended to balance high-level reasoning with cost-effective, high-speed execution for large or complex projects.
Developers can access the model immediately through Anthropic’s API with the identifier `claude-haiku-4-5`. It is also available on Amazon Bedrock and Google Cloud Vertex AI, though Anthropic notes that regional coverage and model IDs within these cloud platforms may update over time. The API price is $1 per million input tokens and $5 per million output tokens. For workflows utilizing prompt-caching, the rates are listed at $1.25 per million write tokens and $0.10 per million read tokens, offering further cost reductions for repeated queries.
Anthropic published benchmark results with methodology details, stating the data shows coding parity with Sonnet 4 and superior performance on computer-use tasks under its test scaffolds. The results include:
- SWE-bench Verified: 73.3% success over 50 trials, using a simple scaffold with two tools (bash, file edits), a 128K thinking budget, default sampling, and no test-time compute.
- Terminal-Bench: An average over 11 runs with the Terminus-2 agent, where six runs had no thinking budget and five had a 32K thinking budget.
- OSWorld-Verified: An average across four runs with a 100-step maximum, a 128K total thinking budget, and a 2K per-step configuration.
- AIME / MMMLU: Averages from multiple runs using default sampling and 128K thinking budgets.
The company advises that users should replicate tests using their own specific orchestration, tool stacks, and thinking budgets before generalizing the performance outcomes to their own applications.
Claude Haiku 4.5 is released under the ASL-2 license. According to Anthropic’s internal tests, the model has a lower measured misalignment rate than both Sonnet 4.5 and Opus 4.1, indicating a different safety profile within the company’s evaluation framework.
