Claude Haiku 5.5 for MCP: Specs & Tool Calling — Test Free

What Claude Haiku 5.5 is

Claude Haiku 5.5 succeeds Haiku 4.5. Anthropic calls it the fastest and most capable small model it has released, aimed at high-volume tasks such as summaries, compaction, database queries and classification, and at sub-agent work alongside Opus 5.5 and Sonnet 5.5. It has a 1M-token window, up from 200K, up to 128K tokens of output, and takes text, image and file input. It is the first Haiku-class model with an adjustable effort setting, from low to max. On Anthropic’s API it is claude-haiku-5-5.

What Anthropic says

OSWorld 2.1 (offline subset): 72.4%, against 15.7% for Haiku 4.5 and 83.9% for Sonnet 5.5.
Terminal-Bench 4.0: 39.2%, against 0.0% for Haiku 4.5 and 70.6% for Sonnet 5.5.
Humanity’s Last Exam with tools: 57.4%, against 18.7% for Haiku 4.5.
GDPval-AA v2.1: 1620, against 735 for Haiku 4.5, 1437 for GPT-6 Luna and 1840 for Sonnet 5.5.
FrontierCode 1.1: 46.4%, against 42.4% for GPT-6 Luna.

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

Natively. Anthropic’s Messages API can connect to remote MCP servers itself. On the API the model id is claude-haiku-5-5.
Through Claude Code, Claude Desktop and claude.ai, all of which are MCP clients
MCP Playground’s Agent Studio: paste your server URL above and run Haiku 5.5 against it in the browser, with no Anthropic key

What to watch for

The effort setting is new for a Haiku, and Anthropic’s launch charts compare results across effort levels. A score read at high effort is not what a client running at low effort gets, so test at the effort your users will have.
OpenRouter lists no temperature, top_p or seed for Haiku 5.5, where Haiku 4.5 had temperature. You cannot narrow run-to-run variation that way; use structured_outputs if tool arguments must come back in a fixed shape.
Terminal-Bench 4.0 is 39.2% against Sonnet 5.5’s 70.6%, and Anthropic itself places Haiku 5.5 at sub-agent and compaction work rather than complex agentic coding. Long, multi-step tool chains are where to check it against Sonnet.
The window is now 1M, but a sub-agent still inherits every tool its parent passes down. Give it the few tools the subtask needs; a smaller model picks the wrong one more often when offered many similar tools.

Run the same four-step task on Haiku 5.5, Haiku 4.5 and Sonnet 5.5 against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. If Haiku 5.5 makes the same calls as Sonnet with correct arguments, it is the cheaper model to try for that server. All the mock servers →

Available here

Model	Model ID
Claude Haiku 5.5	`anthropic/claude-haiku-5.5`

Go deeper

Frequently asked questions

Does Claude Haiku 5.5 support MCP?

Yes, natively. Anthropic’s API can connect to remote MCP servers directly, and Claude Code, Claude Desktop and claude.ai are all MCP clients. You can also run it against any server here without an Anthropic key.

What changed from Claude Haiku 4.5?

The window grows from 200K to 1M tokens, it gains an adjustable effort setting, and Anthropic reports large jumps on agentic benchmarks: OSWorld 2.1 from 15.7% to 72.4%, and Terminal-Bench 4.0 from 0.0% to 39.2%. Both models are available here if you want to compare them on your own tools.

Claude Haiku 5.5 or Sonnet 5.5 for MCP tools?

Sonnet 5.5 is stronger on long agentic work, most clearly on Terminal-Bench 4.0. Haiku 5.5 is faster and suits narrow servers and sub-agent tasks. Run the same prompt on both against your server here; if the calls match, Haiku is the cheaper one to ship.

What is the context window of Claude Haiku 5.5?

1M tokens, as listed on OpenRouter, with up to 128K tokens of output. Tool definitions and every tool result count against it.

Claude Haiku 5.5