Claude Haiku 4.5 Economy
Claude Haiku 4.5 is Anthropic’s fastest and most efficient model, delivering near-frontier intelligence with significantly lower cost and latency than larger Claude models. With performance comparable to Claude Sonnet 4 across reasoning, coding, and computer-use tasks, it brings advanced capabilities to real-time, high-volume applications. As the first Haiku model to support extended thinking, Haiku 4.5 offers adjustable reasoning depth, summarized or interleaved thinking, and tool-assisted workflows spanning coding, Bash, web search, and computer use. Scoring over 73% on SWE-bench Verified, it ranks among the world’s leading coding models while remaining highly responsive for sub-agent orchestration, parallel execution, and large-scale deployment.
Related Models
Claude Fable 5.1 Economy
Claude Fable 5.1 is Anthropic's upgraded flagship frontier model engineered for high-complexity software development, long-horizon agentic workflows, and multi-step knowledge work. Built on Mythos-tier reasoning, Fable 5.1 significantly enhances autonomous self-verification and root-cause troubleshooting while reducing safety false-positive interventions by up to 60%. Powered by an optimized prompt-caching architecture featuring a 75% reduction in cache-read pricing, it cuts total operating costs by 25% to 45% for highly agentic tasks.
Claude Opus 5 Economy
Claude Opus 5 is Anthropic's ultra-flagship frontier model engineered for extreme cognitive density, deep logical reasoning, and expert-level scientific synthesis. Powered by native adaptive thinking and a 1-million-token context window, it resolves complex multi-step mathematical, architectural, and coding challenges with unprecedented accuracy and near-zero hallucination rates. Coupled with rigorous enterprise safety alignment, Claude Opus 5 excels at high-stakes operations, including legal compliance, quantitative financial modeling, advanced biosecurity research, and mission-critical software architecture auditing.
Claude Opus 4.8 Economy
Claude Opus 4.8 is Anthropic’s most powerful officially released Opus model, designed for complex reasoning, long-horizon agentic coding, and highly autonomous professional workflows. It supports text, image, and file inputs and produces text output. It features a 1-million-token context window, up to 128,000 output tokens, and built-in capabilities for adaptive thinking, tool use, and structured outputs. The model excels at advanced coding, browser and computer-use agents, enterprise knowledge work, financial and legal analysis, and multi-step tasks requiring sustained judgment and high reliability.
Claude Fable 5 Economy
Claude Fable 5 is Anthropic’s officially released Mythos-tier model, designed for autonomous knowledge work, advanced reasoning, and long-horizon coding. It supports text, image, and file inputs and produces text output. It features a 1-million-token context window, up to 128,000 output tokens, and built-in always-on adaptive thinking, tool use, and structured outputs. The model shares the same underlying model family as Claude Mythos 5 but was released publicly with stronger safety safeguards. It is suitable for complex software engineering, multi-step agentic workflows, enterprise research, document analysis, and professional tasks requiring sustained reasoning and high reliability.
Read Me
Claude Haiku 4.5 API
Claude Haiku 4.5 is Anthropic's fastest Claude model from the Claude 4.5 generation, built for applications that need near-frontier intelligence without the latency profile of a larger frontier model. It is designed for responsive chat, customer support, pair programming, high-volume processing, and focused tasks inside multi-agent systems.
On iCreat, claude-haiku-4-5 is available through an Anthropic-compatible Messages API. You can test prompts in the Playground, use text and image inputs, enable streaming, and process thinking and text content blocks with familiar Anthropic SDK patterns.
Model Positioning
Claude Haiku 4.5 prioritizes speed and deployment efficiency while retaining strong coding, reasoning, and computer-use capabilities at the model level. Anthropic reported coding performance close to Claude Sonnet 4 at launch, with more than twice the speed and a lower cost on its own platform. These vendor-reported figures establish the model's intended position; production selection should still be based on iCreat tests using your prompts, tools, and latency targets.
Choose Haiku 4.5 when response time and request volume matter, but the workload still needs more than basic text generation. Use a larger Claude model when the hardest reasoning, longest autonomous trajectories, or maximum accuracy matter more than speed.
Key Capabilities
Fast, Near-Frontier Responses
Haiku 4.5 is designed for interactive workloads where waiting time directly affects the user experience. It can support real-time assistants, support queues, coding feedback, and routing systems that must make many decisions quickly.
Coding and Agent Subtasks
The model can handle code generation, debugging, repository questions, and focused execution steps within an agent workflow. Anthropic specifically positions it for pair programming and for parallel subagents coordinated by a more capable model.
Manual Extended Thinking
Extended Thinking can be enabled per request when a problem needs additional reasoning. Haiku 4.5 uses a manually assigned budget_tokens value rather than Adaptive Thinking, allowing applications to reserve deeper reasoning for selected requests instead of applying it to every call.
Long Inputs and Visual Context
The model supports a 200K-token context window, a maximum output of 64K tokens, and text or image input. It can work with long instructions, documents, screenshots, diagrams, and other visual evidence while returning text.
Structured Production Responses
Claude Haiku 4.5 supports structured outputs at the model level, making it suitable for extraction, classification, routing, and application workflows that require responses to follow a defined JSON schema.
Pricing
| Token Type | Price |
|---|---|
| Input | $1 per 1M tokens |
| Output | $5 per 1M tokens |
| Cache Read | $0.10 per 1M tokens |
| Cache Write | $2 per 1M tokens |
Use Cases
- Real-time customer support: classify requests, retrieve relevant context, and draft responsive answers at high request volumes.
- Pair programming: explain code, propose focused changes, diagnose errors, and give fast feedback during development.
- Multi-agent execution: assign parallel research, coding, or verification subtasks while a stronger coordinator handles overall planning.
- Document and image processing: extract, classify, or reason over long text, screenshots, charts, and visual records.
- Latency-sensitive automation: use fast model decisions for routing, moderation support, data transformation, and structured workflows.
Applications that call external tools still need an execution layer, scoped permissions, state management, validation, and stopping rules.
Model Comparison
Claude Haiku 4.5 vs Claude 3.5 Haiku
| Factor | Claude Haiku 4.5 | Claude 3.5 Haiku |
|---|---|---|
| Context Window | 200K tokens | 200K tokens |
| Maximum Output | 64K tokens | 8K tokens |
| Extended Thinking | Supported | Not supported |
| Lifecycle | Current Haiku 4.5 model | Retired on the Anthropic API |
| Best Fit | Fast coding, reasoning, and agent subtasks | Legacy Haiku integrations awaiting migration |
Claude Haiku 4.5 vs GPT-5 and Gemini 2.5 Pro
| Factor | Claude Haiku 4.5 | GPT-5 | Gemini 2.5 Pro |
|---|---|---|---|
| Positioning | Fast, near-frontier Claude model | Frontier coding and agentic reasoning model | Advanced reasoning model for complex tasks |
| Context Window | 200K tokens | 400K tokens | 1,048,576 tokens |
| Maximum Output | 64K tokens | 128K tokens | 65,536 tokens |
| Official Input Modalities | Text and image | Text and image | Text, image, audio, video, and PDF |
| Reasoning Control | Manual Extended Thinking | minimal, low, medium, or high reasoning effort |
Dynamic thinking with configurable thinking budget |
| Best Fit | Low-latency, high-volume Claude workloads | Deep coding and agent workflows | Very long, multimodal reasoning workloads |
Why Choose Claude Haiku 4.5?
Choose Claude Haiku 4.5 when a product needs fast responses without dropping to a basic generation model. Its combination of near-frontier capability, 200K context, image understanding, structured outputs, and optional Extended Thinking makes it a strong default for interactive Claude applications and high-volume task execution.
The practical advantage is routing flexibility: ordinary requests can stay fast, while selected difficult prompts receive a manual thinking budget. Teams can also use Haiku 4.5 for parallel subtasks and reserve larger models for orchestration or the most consequential decisions.
Specifications
| Category | Description |
|---|---|
| Model Name | Claude Haiku 4.5 |
| Provider | Anthropic |
| Model ID | claude-haiku-4-5 |
| Official Snapshot ID | claude-haiku-4-5-20251001 |
| Release Date | October 15, 2025 |
| Model Type | Fast hybrid-reasoning LLM |
| Context Window | 200K tokens |
| Maximum Output | 64K tokens |
| Reliable Knowledge Cutoff | February 2025 |
| Training Data Cutoff | July 2025 |
| Official Input Modalities | Text and image |
| Output Modalities | Text |
| Official Reasoning Control | Manual Extended Thinking with budget_tokens |
| Adaptive Thinking | Not supported |
| iCreat API Features | Anthropic-compatible Messages API, text and image input, streaming, Extended Thinking, thinking/text content blocks |
| Best Suited For | Real-time assistants, high-volume support, pair programming, agent subtasks, structured processing |
Architecture
Anthropic does not publish Claude Haiku 4.5's parameter count or underlying network architecture. The externally documented design is a compact hybrid-reasoning model: it can answer directly for speed or use manually enabled Extended Thinking for more difficult requests.
Extended Thinking does not expose raw chain-of-thought. Applications should treat any returned thinking content as model-generated response metadata and validate the final result with tests, tools, schemas, or trusted evidence.
Production Notes
Enable Extended Thinking only where the expected quality gain justifies additional latency and output-token use. budget_tokens must be lower than max_tokens, and the current turn's thinking tokens count within that output limit.
Haiku 4.5 does not support Adaptive Thinking or Interleaved Thinking. The model cannot alternate hidden reasoning with tool calls in the same way as models that support interleaved thinking; design tool loops and evaluation cases around this limitation.
When migrating from Claude 3.5 Haiku, update the model ID, test prompts again, and handle the refusal stop reason. Claude 4 models accept either temperature or top_p in one request, not both, and older tool versions may require updates.
For multi-turn tool use, preserve the assistant content blocks returned by the API when sending tool results back. Stable prompt prefixes can improve cache reuse, while changes to thinking configuration can affect message cache breakpoints.
FAQ
Is Claude Haiku 4.5 a suitable default for a high-volume application?
It is a strong candidate when response speed, throughput, and capable text or image reasoning must be balanced. Benchmark it on your actual request mix, especially the hardest cases, and route tasks that exceed its quality target to a larger model.
Can Claude Haiku 4.5 replace Claude 3.5 Haiku without retesting the integration?
No. The model ID must change, output behavior can be more concise and direct, and Claude 4 request rules differ in areas such as sampling parameters, refusal handling, and tool versions. Run regression tests before moving production traffic.
Does Extended Thinking run on every Claude Haiku 4.5 request?
No. It must be enabled manually with a thinking-token budget. Use it selectively because thinking consumes part of the output budget and can increase latency and usage.
Can Claude Haiku 4.5 reason between successive tool calls?
It does not support Interleaved Thinking. If an agent depends on reasoning between tool results, keep the loop explicit in application state, preserve returned content blocks, and evaluate whether a model with interleaved thinking is a better fit.