What is Claude Haiku 4.5?
Claude Haiku 4.5 is Anthropic's October 15, 2025 small model, built for speed and volume: near-frontier coding and agent results at the lowest price point in the Claude line. It runs on Unifically as anthropic/claude-haiku-4-5, accepts text and image input, and outputs text. Anthropic's own framing is blunt: it matches Claude Sonnet 4, the frontier model five months before it, on coding, at one-third the cost and more than twice the speed. It is also the first Haiku with extended thinking. Use it where latency and per-call cost decide the design.
Key features of Claude Haiku 4.5
73.3% SWE-bench Verified
Near-frontier agentic coding from a small model. It lands within four points of Sonnet 4.5 and ahead of Sonnet 4, GPT-5, and Gemini 2.5 Pro.
Fastest model in the Claude line
More than twice the speed of Sonnet 4 at one-third the cost, and up to 4-5x the speed of Sonnet 4.5. Fast enough for user-facing chat and in-editor loops.
Extended thinking
The first Haiku that can reason before it answers. Set a thinking budget on hard turns; skip it to keep replies instant.
Computer use and tool calling
It scores 50.7% on OSWorld, ahead of Sonnet 4 at 42.2%, and holds 83.2% on τ2-bench retail. Browser and tool-calling agents get cheap to run at volume.
200k context, 64k output
Room for a full working set of files, a long support thread, or a document batch, with a callable max output of 64k tokens.
Vision input
Text and image input in one request. It reads screenshots, documents, and charts, scoring 73.2% on MMMU.
Best for
Real-time chat assistants
Latency low enough to keep a typed conversation moving.
Customer support agents
Tool calls, ticket context, and fast replies at high volume.
Pair programming
In-editor suggestions and fixes without the wait of a larger model.
Subagent fleets
A Sonnet or Opus orchestrator plans; a team of Haiku workers runs the subtasks in parallel.
High-volume pipelines
Classification, extraction, and summaries where per-call cost adds up fast.
Budget computer use
Browser tasks that click, type, and read screens, at small-model prices.
Use cases
Build a coding assistant that lives in the editor: with 73.3% on SWE-bench Verified and low latency, Claude Haiku 4.5 handles the fix-test-commit loop without breaking flow. Use it as the worker model in a multi-agent system, where Sonnet 4.5 breaks a problem into steps and a team of Haiku 4.5 instances completes subtasks in parallel. It also fits support automation that reads a ticket, calls tools, and answers in seconds, and document pipelines that classify, extract, and summarize thousands of items a day. Vision input covers screenshot triage and chart reading in the same call.
Limitations
Claude Haiku 4.5 trails the bigger Claude models on deep reasoning. It scores 73.0% on GPQA Diamond versus 83.4% for Sonnet 4.5, and 50.7% on OSWorld versus 61.4%.
The context window is 200k tokens, not the 1M of newer Sonnet and Opus models, and max output is 64k tokens. Whole-repo work at the largest scale belongs on a bigger model.
Its reliable knowledge cutoff is February 2025. Pair it with search or retrieval when recent facts matter.
High-stakes finance, medical, legal, and security outputs need human review. That rule does not relax because the model is fast.
Claude Haiku 4.5 vs Claude Sonnet 4.5
Sonnet 4.5 is the stronger model: 77.2% versus 73.3% on SWE-bench Verified, 61.4% versus 50.7% on OSWorld, and 83.4% versus 73.0% on GPQA Diamond. Haiku 4.5 is the faster, cheaper one, with up to 4-5x the speed at a fraction of the cost. On tool calling the gap nearly closes: 83.2% versus 86.2% on τ2-bench retail. Pick Sonnet 4.5 when one agent must carry a long, hard task alone. Pick Haiku 4.5 when the work splits into many fast calls, or when a user is watching the cursor blink.
When to use Claude Haiku 4.5
Use Claude Haiku 4.5 when speed and per-call cost decide the design: user-facing chat, support automation, in-editor coding help, high-volume pipelines, and subagent workers under a larger orchestrator. Move up to a Sonnet or Opus model when a single agent needs the deepest reasoning or a 1M-token context.
API examples
Call Claude Haiku 4.5 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/anthropic/claude-haiku-4-5.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "anthropic/claude-haiku-4-5",
"messages": [
{ "role": "user", "content": "Classify this support ticket and draft a first reply." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
anthropic/claude-haiku-4-5, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
200k tokens, with a callable max output of 64k tokens. That fits a working set of source files, a long support thread, or a batch of documents in one request.
Yes for the price. It scores 73.3% on SWE-bench Verified, close to Sonnet 4.5 and ahead of Sonnet 4, with far lower latency and cost per call.
Yes. It is the first Haiku with extended thinking, so you can set a thinking budget on hard turns and keep instant replies on easy ones.
Text and image input with text output. It reads screenshots, documents, and charts in the same request as code or chat.
It is the fastest model in the Claude line. Anthropic measured more than twice the speed of Sonnet 4 at one-third the cost, and up to 4-5x the speed of Sonnet 4.5.
Deep reasoning trails the bigger Claude models, the reliable knowledge cutoff is February 2025, and high-stakes finance, legal, medical, and security outputs need human review.
