Claude Opus 5.5

Claude Opus 5.5 is Anthropic’s Opus model for long-running agentic coding and knowledge work. Use the AnyFast model ID claude-opus-5-5 with the native Anthropic Messages API at POST /v1/messages.

Model specifications

Specification	Value
Model ID	`claude-opus-5-5`
Context window	1M tokens
Maximum output	128K tokens
Input → output	Text and images → text
Thinking	Adaptive, always on
Default effort	`medium`
Effort levels	`low`, `medium`, `high`, `xhigh`, `max`
Reliable knowledge cutoff	June 2026

What changed from Opus 5

Always-on thinking — Omit thinking or set it to {"type":"adaptive"}. Both disabled and manual enabled modes return 400.
New default effort — Requests that omit output_config.effort use medium, rather than the high default used by Opus 5.
No forced tool use — tool_choice supports auto and none. The any and named tool modes return 400.
Preserved thinking — Keep the conversation prefix unchanged and return prior thinking blocks exactly as received during tool-use and multi-turn workflows.
Refusal handling — A safeguard refusal can return HTTP 200 with stop_reason: "refusal" and additional information in stop_details.

Quick example

curl https://www.anyfast.ai/v1/messages \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 4096,
    "output_config": { "effort": "medium" },
    "messages": [
      {
        "role": "user",
        "content": "Review this migration plan and identify the three highest-risk assumptions."
      }
    ]
  }'

import anthropic

client = anthropic.Anthropic(
    api_key="YOUR_API_KEY",
    base_url="https://www.anyfast.ai"
)

message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    output_config={"effort": "medium"},
    messages=[
        {
            "role": "user",
            "content": "Review this migration plan and identify the three highest-risk assumptions."
        }
    ]
)

for block in message.content:
    if block.type == "text":
        print(block.text)

import anthropic

client = anthropic.Anthropic(
    api_key="YOUR_API_KEY",
    base_url="https://www.anyfast.ai"
)

with client.messages.stream(
    model="claude-opus-5-5",
    max_tokens=4096,
    output_config={"effort": "medium"},
    messages=[
        {"role": "user", "content": "Create a phased rollout plan for this database migration."}
    ],
) as stream:
    for text in stream.text_stream:
        print(text, end="")

Core parameters

Parameter	Type	Required	Description
`model`	string	Yes	Must be `claude-opus-5-5`.
`messages`	array	Yes	Conversation turns with `user` or `assistant` roles. Text, images, and tool results are supported.
`max_tokens`	integer	Yes	Maximum output tokens across thinking and visible response text. Range: 1–128000.
`output_config.effort`	string	No	Controls thinking depth. Supports `low`, `medium`, `high`, `xhigh`, and `max`; default is `medium`.
`thinking.type`	string	No	Omit it or set it to `adaptive`. Thinking cannot be disabled.
`tools`	array	No	Tool definitions available to the model.
`tool_choice.type`	string	No	Supports `auto` and `none`. Forced `any` or a named `tool` is not supported.
`stream`	boolean	No	Enables SSE streaming. Default: `false`.
`stop_sequences`	array	No	Custom sequences that stop generation.

Adaptive thinking is always on. Use output_config.effort to balance reasoning depth, latency, and token use. Leave enough room in max_tokens for both thinking and visible output, especially at xhigh or max effort.

Do not send thinking: {"type":"disabled"} or manual extended thinking with budget_tokens; both return 400. Keep tool_choice at auto or none because forced any and named tool choices are also rejected.

Tool use

Keep tool_choice at auto (the default) or use none. Describe each tool and its input schema precisely, then state in the prompt when the tool should be used. In a tool loop, inspect content blocks by type instead of assuming the first block contains text.

Multi-turn conversations

When an assistant turn contains thinking blocks, send the complete turn back exactly as received. Treat the system prompt, tool definitions, and earlier messages as append-only. Changing content before an Opus 5.5 thinking block can invalidate the block and return 400.

Migrating from Opus 5

Change the model ID from claude-opus-5 to claude-opus-5-5.
Remove thinking: {"type":"disabled"} and manual budget_tokens settings.
Set output_config.effort explicitly if you need behavior comparable across model versions.
Replace forced tool_choice modes with auto or none.
Update refusal handling to check stop_reason and stop_details, even when the HTTP status is 200.

Anthropic released Claude Opus 5.5 on September 22, 2026. See the official model overview, what’s new, and announcement.

API Reference

View request fields, response blocks, and the interactive API playground.

Getting started

Model Introduction

Resources

Claude Opus 5.5

Model specifications

What changed from Opus 5

Quick example

Core parameters

Tool use

Multi-turn conversations

Migrating from Opus 5

API Reference

​Model specifications

​What changed from Opus 5

​Quick example

​Core parameters

​Tool use

​Multi-turn conversations

​Migrating from Opus 5

API Reference

Model specifications

What changed from Opus 5

Quick example

Core parameters

Tool use

Multi-turn conversations

Migrating from Opus 5