Anthropic’s newest small model, Claude Haiku 4.5, delivers similar coding performance to earlier frontier models at dramatically lower cost and latency—ideal for real-time chat, customer support, pair programming, and multi-agent systems.
What Is Claude?
Claude is Anthropic’s family of large language models focused on reliability and safety. The lineup includes Opus (frontier intelligence), Sonnet (balanced power), and Haiku (lightweight, ultra-fast, cost-efficient). The new Claude Haiku 4.5 pushes speed and efficiency while preserving high accuracy for coding and tool-use tasks.
What’s New in Claude Haiku 4.5
- Near-frontier coding performance—similar to Claude Sonnet 4 on real-world tasks (SWE-bench Verified).
- Much faster responses—more than twice the speed of Sonnet 4 for low-latency workflows.
- Lower cost—about one-third the cost of Sonnet (pricing below).
- Better “computer use” and tool use—surpasses Sonnet 4 on OSWorld and strong scores on τ2-bench.
- Multi-agent orchestration—use Sonnet 4.5 to plan and coordinate a team of Haiku 4.5 workers in parallel.
- Wider availability—Claude apps, Claude Code, API, Amazon Bedrock, and Google Cloud Vertex AI.
Benchmark Highlights
Claude Haiku 4.5 offers roughly 90% of Sonnet 4.5 performance at a fraction of the cost and latency. Key published results include:
Methodology (condensed)
SWE-bench Verified: 73.3% averaged over 50 trials using bash + file-edit tools, no test-time compute, 128K thinking budget, default sampling; includes minor prompt addendum encouraging tool use and writing tests first.
Terminal-Bench: Default Terminus-2 framework, 11 runs (6 without thinking, 5 with 32K thinking), n-attempts=1.
τ2-bench: 10-run average with extended thinking (128K), default sampling, tool use and targeted prompt addenda for Airline/Telecom policies.
AIME: Average over 10 independent runs; pass@1 over 16 trials, default sampling, 128K thinking.
OSWorld: Official OSWorld-Verified framework, 100 max steps, average of 4 runs, 128K total thinking, 2K per step.
MMMLU: Average of 10 runs across 14 non-English languages with 128K thinking. Other scores: 10-run averages with default sampling and 128K thinking.
Pricing & Availability
- Model name:
claude-haiku-4-5 - Price: $1 per million input tokens / $5 per million output tokens
- Where to use: Claude apps & Claude Code, Claude API, Amazon Bedrock, and Google Cloud Vertex AI
- Migration: Drop-in replacement for Haiku 3.5 and Sonnet 4 in many workflows
Safety & Alignment
Haiku 4.5 showed low rates of concerning behavior in Anthropic’s evaluations and a statistically lower misalignment rate than Sonnet 4.5 and Opus 4.1, making it the safest Claude model by that metric. It poses only limited CBRN risk and is released under AI Safety Level 2 (ASL-2). See the model’s system card for details.
How Teams Are Using Haiku 4.5
- Real-time chat & support: low latency for conversational agents and customer service workflows.
- Pair programming & Claude Code: faster turnarounds for prototyping and code generation.
- Multi-agent pipelines: use Sonnet 4.5 for planning and orchestrate many Haiku 4.5 workers to run subtasks in parallel.
- High-volume inference: scale to millions of calls while controlling cost.
FAQ
Is Claude Haiku 4.5 better than Sonnet 4.5?
Sonnet 4.5 remains Anthropic’s frontier model and the best overall coding model. Haiku 4.5 aims for similar coding and tool-use performance at much lower cost and latency, making it ideal for real-time, large-scale deployments.
How fast and how affordable is Haiku 4.5?
Haiku 4.5 delivers more than 2× the speed of Sonnet 4 and costs roughly one-third as much as Sonnet, at $1 per million input tokens and $5 per million output tokens.
Where can I access Claude Haiku 4.5?
Use it in the Claude apps and Claude Code, or via the Claude API (claude-haiku-4-5). It’s also available on Amazon Bedrock and Google Cloud Vertex AI.
Does Haiku 4.5 replace Haiku 3.5?
For many use cases it’s a drop-in upgrade, offering stronger performance while remaining highly economical. Existing Haiku 3.5 and Sonnet 4 users can migrate with minimal changes.
What’s the safety level of Haiku 4.5?
Anthropic classifies the model at AI Safety Level 2 (ASL-2), reflecting limited CBRN risk and lower misalignment rates compared to other Claude 4.5 models.
