Claude Opus 5.5 API — pricing, model ID and what changed from Opus 5

Claude Opus 5.5 is Anthropic's newest Opus, released September 22, 2026, and the model Anthropic now points "most workloads" at. Its API price is $4.00 input / $20.00 output per 1M tokens at Anthropic — below Opus 5's $5.00 / $25.00 — and $1.60 / $8.00 on Kunavo, 60% under that list. The model ID is claude-opus-5-5, it has a 1M-token context window and up to 128K output tokens, and it thinks on every request.

Claude Opus 5.5 API pricing

Anthropic's published rates next to Kunavo's, per 1M tokens. Anthropic reads Opus 5.5's cache at 5% of input rather than the 10% it charges on most Claude models, and Kunavo passes the same ratio through:

Next to the rest of the Claude line on Kunavo, Opus 5.5 is the cheapest way to get Opus-tier reasoning — cheaper than Claude Opus 5 ($2.00 / $10.00) and than Claude Sonnet 5 ($2.00 / $10.00), with Claude Fable 5.1 at $7.00 / $35.00 above it and Claude Haiku 4.5 at $0.40 / $2.00 below. Every Claude rate, with worked examples, is in the Claude API pricing guide; the model's own page is /models/claude-opus-5-5.

Call Claude Opus 5.5

One Kunavo key reaches Opus 5.5 through both the OpenAI-compatible endpoint and the native Anthropic Messages API. With the OpenAI SDK, change base_url and the model string:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["KUNAVO_API_KEY"],
    base_url="https://api.kunavo.com/v1",
)

# Opus 5.5 always thinks, and thinking bills as output — leave max_tokens
# generous or the budget is spent before the answer starts.
resp = client.chat.completions.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Review this migration plan for failure modes."}],
)
print(resp.choices[0].message.content)

With the Anthropic SDK, pass the origin as base_url; the SDK appends /v1/messages itself:

from anthropic import Anthropic

# Native Messages API: pass the origin — the SDK appends /v1/messages.
client = Anthropic(base_url="https://api.kunavo.com", auth_token="sk-kn-...")

msg = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=8192,   # thinking plus the answer
    messages=[{"role": "user", "content": "Find the race condition in this diff."}],
)
for block in msg.content:
    if block.type == "text":
        print(block.text)

Keys are created at /app/keys; the Claude API key guide walks through it.

What changed from Opus 5 — and what it does to your requests

Anthropic lists four breaking changes for code that ran on Opus 5, and one change to the response shape. Write for Anthropic's rules — but know that two of them did not fail through Kunavo when we tested on September 24, 2026:

• Thinking cannot be disabled. At Anthropic, thinking: {type: "disabled"} and budget_tokens return a 400 at every effort level, and effort (default medium) is the only control. Through Kunavo the disable request came back 200 and the model thought anyway, and output_config.effort made no measurable difference — the channel serving Opus 5.5 does not apply either setting. Thinking bills as output: a one-word reply spent 15 of its 16 output tokens thinking. Leave max_tokens generous.
• Forced tool use is not supported. At Anthropic, tool_choice of any or tool returns a 400 (the OpenAI-compatible "required" maps to any). Through Kunavo a forced-tool request returned 200 instead — so the forcing is not something to rely on. Use auto and name the tool in the prompt.
• Thinking blocks are tied to the model and the conversation. Replay them unchanged on Opus 5.5; editing the system prompt, tools or earlier turns invalidates them. See the thinking-signature error guide.
• The older computer-use tool is gone. computer_20251124 is not accepted on the Claude API.
• Text between tool calls arrives as thinking blocks, empty at the default display setting — an app that streams progress text goes quiet between tool calls.

Sampling parameters were already gone on Opus 5; on Opus 5.5 Kunavo strips temperature, top_p and top_k before the request leaves the gateway, as on the rest of the Claude 5 family.

Opus 5.5 vs Opus 5, by Anthropic's numbers

Anthropic's announcement says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5", and that it generates output more than 30% faster. Its published scores against Opus 5 (Anthropic's figures, not Kunavo's):

For new work there is little reason to stay on Opus 5: Opus 5.5 is faster, scores higher and costs less, and the model string is the only change unless your code relies on disabling thinking or on forcing a tool. Stay on claude-opus-5 only until those two paths are rewritten.

Claude Code on Kunavo with Opus 5.5

From Claude Code v2.1.280, its default model and its opus alias resolve to Opus 5.5, and Kunavo serves it — so an unpinned Claude Code on a current version runs on Opus 5.5 here. Kunavo's setup guide pins the opus alias to claude-opus-5, which every Claude Code version can use. To run a session on Opus 5.5, set ANTHROPIC_MODEL:

export ANTHROPIC_BASE_URL=https://api.kunavo.com   # origin, no /v1
export ANTHROPIC_AUTH_TOKEN=sk-kn-...

# Run the session on Opus 5.5.
export ANTHROPIC_MODEL=claude-opus-5-5
# Kunavo's setup guide keeps the opus alias on claude-opus-5.
export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5

claude

The rest of the setup — the origin-only base URL, why ANTHROPIC_AUTH_TOKEN and not ANTHROPIC_API_KEY, and the settings.json form — is on the Claude Code page.

FAQ