Anthropic released Claude Haiku 4.5, a latency-optimized “small” model with coding performance similar to Claude Sonnet 4. The company reports Haiku 4.5 runs more than twice as fast as Sonnet 4 at one-third the cost. It is available immediately through Anthropic’s API and in partner catalogs on Amazon Bedrock and Google Cloud Vertex AI.
Pricing for Haiku 4.5 is set at $1 per million input tokens and $5 per million output tokens. For prompt-caching, the listed rates are $1.25 per million write tokens and $0.10 per million read tokens.
Anthropic positions Haiku 4.5 for workloads with tight latency budgets and high throughput, including real-time assistants, customer-support automations, and pair-programming. The company suggests it 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 involving GUI and browser manipulation, as seen in products like Claude for Chrome. It also improves responsiveness in Claude Code for multi-agent projects and rapid prototyping.
Within Anthropic’s model lineup, Sonnet 4.5 remains the company’s frontier model, which it describes as “the best coding model in the world.” Haiku 4.5 is positioned to offer near-frontier performance with greater cost-efficiency. Anthropic recommends an orchestration pattern where Sonnet 4.5 handles multi-step planning, with a pool of Haiku 4.5 workers managing parallelized execution.
Developers can access the model on Anthropic’s API using the identifier claude-haiku-4-5. While Anthropic confirmed its availability on Amazon Bedrock and Google Cloud Vertex AI at launch, the company notes that specific model catalog IDs and regional coverage may change over time.
Anthropic provided benchmark results to demonstrate the model’s capabilities. On SWE-bench Verified, Haiku 4.5 achieved a 73.3% score using a simple scaffold with bash and file edit tools. This result was averaged over 50 trials with a 128K thinking budget and no test-time compute. The company also shared results from Terminal-Bench, tested with a Terminus-2 agent across 11 runs, and OSWorld-Verified, tested with a 100-step maximum over four runs. Performance on AIME and MMMLU was also measured using default sampling and 128K thinking budgets.
The company stated these results show coding parity with Sonnet 4 and performance gains in computer-use tasks under the specific testing scaffolds. It advises users to replicate tests with their own orchestration and tool stacks before generalizing performance. The model is released under the ASL-2 license. In internal tests, Anthropic reports Haiku 4.5 had a lower measured misalignment rate than both Sonnet 4.5 and Opus 4.1.
