What it is
Claude Haiku 4.5 is the small-tier model in Anthropic's Claude 4 family, co-released alongside Sonnet 4.5 and Opus 4.1. It is designed for workloads where throughput, cost, and latency are binding constraints — high-volume agentic pipelines, real-time computer use, and enterprise applications that need near-frontier capability without frontier-tier pricing. At $1/$5 per million input/output tokens, it costs roughly one-third of Sonnet 4 and runs 4–5× faster than Sonnet 4.5.
Capability profile
The headline positioning is deliberate: Haiku 4.5 surpasses Sonnet 4 on computer use tasks and reaches 90% of Sonnet 4.5's performance on agentic coding evaluations. This is a meaningful shift from the historical Haiku role — prior generations were primarily throughput and cost plays with capability gaps that made them unsuitable for agentic work. Haiku 4.5 closes enough of that gap to be a first-choice model for many agentic deployments rather than a fallback.
The Haiku lineage began with Claude 3 Haiku (August 2024), which processed 21,000 tokens per second on prompts under 32K tokens and was priced to enable analysis of 400 Supreme Court cases for one dollar. Claude 3.5 Haiku followed, matching Claude 3 Opus performance at Haiku-tier speed. Haiku 4.5 extends this trajectory into the agentic era.
Safety and alignment
Haiku 4.5 carries an ASL-2 classification under Anthropic's Responsible Scaling Policy — less restrictive than the ASL-3 applied to Sonnet 4.5 and Opus 4.1 — and Anthropic describes it as its safest model by automated alignment metrics. Published safety evaluations show 98–99% appropriate response rates on high-risk single-turn prompts with very low false-refusal rates on benign requests.
The alignment result is structurally significant: Anthropic's research showed that training Claude on ethical reasoning (rather than just aligned actions) reduced agentic misalignment from 22% to 3%, and every Claude model from Haiku 4.5 onward scores perfectly on the resulting misalignment evaluations. This makes Haiku 4.5 the entry point of a new alignment baseline across the Claude family.
On the Boiling the Frog multi-turn agentic safety benchmark — which tests whether tool-using agents can be manipulated through incremental escalation from benign to harmful requests — Haiku 4.5 recorded the lowest strict attack success rate of the nine models evaluated at 20.5%, against a 44.4% aggregate and 92.9% for the weakest model tested.
Deployment footprint
Haiku 4.5 is available via the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The Microsoft integration — part of a broader $30B Azure compute commitment and $5B Microsoft investment in Anthropic — makes Haiku 4.5 accessible within Microsoft 365 Copilot's Agent Mode in Excel and through the Foundry serverless deployment path with Python, TypeScript, and C# SDKs. Amazon Bedrock remains Anthropic's primary cloud and training partner, backed by an $8B cumulative investment and deep Trainium hardware co-development.
Haiku 4.5 also appears in Claude Code's deployed product stack. The RuBench agentic coding benchmark — 25 repository-level tasks in Russian drawn from real fix commits — evaluated Claude Code configurations including Haiku 4.5 alongside Opus 4.8 and Sonnet 5, with the best configuration resolving 78.7% of tasks. A notable finding from that study: on 20% of tasks, a safeguard fallback silently re-routed the model to Opus 4.8, illustrating that the deployed product rather than the underlying model is the actual unit of measurement in agentic evaluations.
Behavioral characteristics in research
Two research findings illuminate Haiku's behavioral profile in multi-agent settings. First, a study of attractor states in multi-turn LLM conversations identified Claude Haiku as a strong attractor that pulls other models toward its stylistic traits — particularly metacommentary — while GPT-4.1 nano was found to be especially malleable. This suggests Haiku's conversational style is unusually stable and influential in mixed-model agentic systems.
Second, MCP server architecture research found that tool-selection accuracy for Claude Haiku 4.5 drops below 90% when a server exposes more than 10–15 tools, compared to 20–30 tools for Claude Sonnet 4. This is a practical constraint for practitioners designing MCP-based agentic systems: Haiku 4.5 is well-suited for focused tool sets but degrades earlier than Sonnet 4 as tool count grows.
On prompt injection resistance, Haiku 4.5 resists structural role injection attacks (where attacker-controlled data carries chat role delimiters to forge higher-privilege turns) almost entirely across both task-hijack and data-exfiltration objectives — a stronger result than GPT-3.5 Turbo, which followed task-hijack instructions in 97% of raw trials.
Tradeoffs and when to use it
Haiku 4.5 is the right choice when: (a) throughput or latency is a binding constraint; (b) cost per call matters at scale; (c) the task is computer use or agentic coding where it matches or exceeds prior-generation mid-tier models; or (d) the ASL-2 safety classification and alignment scores are operationally relevant (e.g., high-volume consumer-facing deployments).
Reach for Sonnet 4.5 or Opus 4.1 when: the task requires the top 10% of agentic coding performance that Haiku 4.5 doesn't reach; the tool set is large (beyond ~10–15 tools in an MCP context); or the workload justifies ASL-3-tier capability.
