<MCP.Playground/>
online

Claude Haiku 4.5

Runs MCP tools

Anthropic’s small, fast Claude from October 2025, since succeeded by Haiku 5.5. It is the model Anthropic suggested for sub-agents: a larger Claude plans, and several Haiku 4.5 instances carry out the tool calls in parallel.

Vendor
Anthropic
Released
15 October 2025
Context
200K tokens
Input
Text, image, file

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

In your browser

MCP Agent Studio

Server has a public URL? Chat with Claude Haiku 4.5 on it, watch every tool call, and compare models side by side.

Open MCP Agent Studio →

On your computer

MCP Playground Desktop

Server runs on localhost or over stdio? The browser can’t reach it. The desktop app can, with Claude Haiku 4.5 in Agent Studio. Download free

Mac (Apple Silicon)Mac (Intel)Windowsor Microsoft Store

MCP Evals

Can Claude Haiku 4.5 actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Claude Haiku 4.5 drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Claude Haiku 4.5 as the driver.

Run evals →See a sample

What Claude Haiku 4.5 is

Claude Haiku 4.5 is the fast, low-cost tier of the Claude 4.5 generation. Anthropic says it matches Claude Sonnet 4 on coding at one-third of the cost and more than twice the speed, and beats Sonnet 4 at computer use. It supports extended thinking with a budget you set, takes text, image and PDF input, and has a 200K-token window. Anthropic released it under ASL-2, a lower safety level than the ASL-3 of Sonnet 4.5 and Opus 4.1.

What Anthropic says

  • SWE-bench Verified: 73.3%, averaged over 50 trials, against 77.2% for Claude Sonnet 4.5.
  • OSWorld: 41.75%, with a 32K thinking budget.
  • Anthropic positions it as the worker in multi-agent setups, with Sonnet 4.5 splitting the problem and several Haiku 4.5s doing the subtasks.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

  • Claude Code, Claude Desktop and the Claude API, all of which connect to MCP servers directly
  • Anthropic’s MCP connector on the Messages API, which lets the model call a remote MCP server without a client in between
  • MCP Playground’s Agent Studio: paste your server URL above and run Haiku 4.5 against it in the browser, with no Anthropic key

What to watch for

  • Thinking is off unless you ask for it. Anthropic’s benchmark numbers were run with thinking budgets of 32K–128K, so a client that calls Haiku with no budget is testing a different setup from the one in the launch post.
  • The window is 200K, not the 1M of newer models. Servers that return large tool results, such as whole files, logs or query dumps, fill it in a few turns.
  • It is the cheap choice for sub-agents, but a sub-agent inherits every tool its parent passes down. Give it the few tools the subtask needs; a smaller model picks the wrong tool more often when offered many similar ones.

Run the same four-step task on Haiku 4.5 and Claude Sonnet 5 against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. If Haiku gets the arguments right on every step, it is worth trying as the cheaper model for that server. All the mock servers →

Available here

ModelModel ID
Claude Haiku 4.5anthropic/claude-haiku-4.5

Go deeper

Frequently asked questions

Does Claude Haiku 4.5 support MCP?
Yes. Claude Code, Claude Desktop and the Claude API all connect it to MCP servers, and Anthropic’s MCP connector can reach a remote server straight from the Messages API. You can also run it against your server in MCP Playground.
Claude Haiku 4.5 or Claude Sonnet 5 for MCP tools?
Sonnet 5 is the stronger model with a newer training run. Haiku 4.5 is faster and cheaper per call. For narrow servers with a few clear tools Haiku is often enough; for long multi-step chains, test both here first.
Does Claude Haiku 4.5 have extended thinking?
Yes, and it is the first Haiku model that does. You turn it on per request with a thinking budget. Without one, Haiku answers directly.
What is the context window of Claude Haiku 4.5?
200K tokens, as listed on OpenRouter. Tool definitions and every tool result count against it.

Sources