Introduction
Claude Opus 5.5 is Anthropic's newest Opus, released on 22 September 2026 and served here on one OpenAI-compatible endpoint at /v1/chat/completions. A 1,000,000-token window carries the whole prior exchange verbatim, and one reply can run to 128,000 tokens.
Thinking is always on. reasoning_effort takes low, medium or high, with medium as the default. The model fixes its own sampling, so temperature and top_p are not accepted; message content is text and images, up to 128 tools can be defined per request, tool_choice takes auto or none, and replies stream over Server-Sent Events.
Key features & highlights
Memory measured in turns, not paragraphs
Thirty turns in, does the character still know where she came from, what she promised and what she said she would never do? With 1,000,000 tokens of input the answer depends only on what you send: put the full history in the request and the model reads the original lines, not a paraphrase of them.
Faster than Opus 5, by Anthropic's figures
Anthropic reports that Opus 5.5 writes more than 30% faster than Opus 5 and, at default settings, uses fewer tokens on typical workloads. On a long streamed reply, that is time the reader does not spend waiting.
Always-on thinking, three effort levels
Opus 5.5 always reasons before it answers, and that reasoning is billed as output. reasoning_effort decides how much: low for quick conversational turns, medium for most work, high when planted details and established facts must be reconciled first.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | Ordered conversation turns. Content is a string, or an array of text and |
max_tokens | No | Caps the reply at 128,000 - an eighth of the read window. Default 4,096; raise it before writing a chapter. |
reasoning_effort | No | low, medium or high, medium by default - the same default Anthropic sets for this model. Thinking cannot be switched off on Opus 5.5; this only sets how much of it happens. |
stop | No | Up to 4 sequences, as a string or an array. The reply ends as soon as the model writes one of them, and the matched text itself is left out. at least 1 characters · at least 1 items · at most 4 items |
tools | No | Up to 128 function definitions per request. at most 128 items |
parallel_tool_calls | No | true or false, default true: the model may request several tool calls in one turn. Default: |
tool_choice | No | auto or none. Opus 5.5 does not accept a forced tool: to get a specific call, name the tool in the prompt and leave this on auto. |
Pricing
Input $3.60 / Output $18 per 1 million tokens
Use cases
A persona held across dozens of turns
Send the conversation as written and the character still knows, thirty turns later, where she came from, what she agreed to and what she refuses to do. Most long exchanges need no summarisation layer at all.
Agents on long task chains
Hand over the repository, the specification and the discussion so far in one request and let the model decide what to read first. The faculty a long roleplay tests is the one a multi-step engineering task needs: at step ten, does it still remember the decision from step one?
Structured extraction through tools
Define a tool whose parameters are the fields you need and ask the model to answer by calling it. With the whole source in the window, a missing field points to a gap in the material rather than in the pipeline.
Prompt guide & pro tips
Send the whole history, oldest first
The window is usually larger than the summariser someone was about to build. Put every earlier message into the messages array in order and let the model read what was actually said.
Persona, setting and limits in the system message
The system message is the one slot that never scrolls away in a long exchange. Origin, speech habits, hard limits, narrative person and tense belong there; the user message carries only the current turn.
Ask for a tool call instead of forcing one
Opus 5.5 rejects a forced tool_choice, and there is no response_format on this endpoint. Define a tool whose parameters are the JSON schema you want, leave tool_choice on auto, tell the model to answer by calling that tool, and validate the arguments. A conversation cannot end on an assistant message to prefill the reply, and max_tokens above 128,000 and video and audio parts are rejected.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "anthropic/claude-opus-5.5/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.
