Anthropic has introduced Claude Haiku 4.5, a compact, latency-optimized AI model that matches the coding capabilities of Claude Sonnet 4 while operating at over twice the speed and costing only a third as much. This model is immediately accessible through Anthropic’s API and is also integrated into partner platforms such as Amazon Bedrock and Google Cloud Vertex AI. The pricing structure is set at $1 per million input tokens and $5 per million output tokens. Designed as a seamless upgrade, Haiku 4.5 is positioned to replace both Haiku 3.5 and Sonnet 4 in scenarios where cost efficiency and low latency are critical.
Optimized for Real-Time and Interactive Applications
Haiku 4.5 is tailored for environments demanding rapid response times and high throughput, such as live virtual assistants, automated customer support, and collaborative coding sessions. It notably outperforms Sonnet 4 in tasks involving graphical user interface and browser interactions-key components behind tools like Claude for Chrome. This enhancement significantly boosts the agility of Claude Code in multi-agent workflows and fast-paced prototyping. While Anthropic maintains that Sonnet 4.5 remains the pinnacle for coding tasks, branding it as “the world’s best coding model,” Haiku 4.5 offers nearly equivalent performance with superior cost-effectiveness. A recommended deployment strategy involves leveraging Sonnet 4.5 for complex, multi-step planning, complemented by a fleet of Haiku 4.5 instances executing tasks in parallel.
Access, Model Identification, and Cost Details
Developers can immediately access the model via Anthropic’s API under the identifier claude-haiku-4-5. Additionally, Haiku 4.5 is available through major cloud marketplaces including Amazon Bedrock and Google Cloud Vertex AI, with regional availability and model IDs subject to updates. The API pricing is competitive, charging $1 per million tokens for inputs and $5 per million tokens for outputs. Anthropic also offers prompt caching services priced at $1.25 per million tokens written and $0.10 per million tokens read, enabling cost savings for repeated queries.
Performance Benchmarks and Evaluation
Anthropic has provided comprehensive benchmark results across various standardized and agent-based testing suites, detailing the evaluation methodologies to ensure transparency:
- SWE-bench Verified: Utilizes a simple framework with two tools (bash commands and file editing), achieving an average success rate of 73.3% over 50 trials without additional compute during testing, operating under a 128K token thinking budget. The prompt encourages extensive tool usage and prioritizes writing tests first.
- Terminal-Bench: Employs the Terminus-2 agent, averaging results from 11 runs-6 without additional thinking and 5 with a 32K token thinking budget.
- OSWorld-Verified: Conducted with a maximum of 100 steps, averaged over 4 runs using a 128K total thinking budget and 2K tokens per step.
- AIME / MMMLU: Aggregates multiple runs with default sampling and a 128K token thinking budget.
These benchmarks highlight Haiku 4.5’s parity with Sonnet 4 in coding tasks and its superior performance in computer-use scenarios. Users are encouraged to validate these results within their own environments, considering their specific orchestration, toolsets, and computational budgets.
Summary of Advantages
- Haiku 4.5 achieves coding performance on par with Sonnet 4 while reducing costs by two-thirds and doubling processing speed.
- It excels beyond Sonnet 4 in tasks involving computer interface manipulation, enhancing responsiveness in applications like Claude for Chrome and multi-agent coding workflows.
- Optimal usage involves Sonnet 4.5 handling complex planning stages, with Haiku 4.5 managing parallel execution tasks.
- Available through multiple platforms with transparent pricing: $1 per million input tokens and $5 per million output tokens.
- Released under the permissive ASL-2 license, Haiku 4.5 demonstrates a lower misalignment rate compared to Sonnet 4.5 and Opus 4.1, supporting safer and more reliable enterprise deployment.
Analysis and Industry Implications
Anthropic’s launch of Claude Haiku 4.5 represents a strategic advancement in AI model deployment, balancing high coding proficiency with operational efficiency. By offering a model that delivers near-top-tier performance at a fraction of the cost and with significantly reduced latency, Anthropic empowers developers to architect systems with a clear division of labor: Sonnet 4.5 for intricate planning and Haiku 4.5 for scalable execution. This approach minimizes the need for extensive architectural overhauls, as Haiku 4.5 serves as a direct substitute across APIs and cloud marketplaces. Furthermore, the ASL-2 licensing and documented improvements in alignment safety lower barriers for adoption in enterprise environments where compliance and cost control are paramount.
