Claude Opus 5.5

Claude Opus 5.5 is Anthropic's first model in the Claude 5.5 family, released on 22 September 2026. It improves on Opus 5 in coding, computer use and knowledge work, while reducing the price per token.

Anthropic claims it approaches Claude Fable 5.1 on most work, with lower running costs (fewer tokens per task) - apparently 40% cheaper than Opus 5 on typical workloads at default settings.

Availability and context

As of 23 September 2026, Opus 5.5 is available through Claude, the Claude API, AWS, Google Cloud and Microsoft Azure. The direct API identifier is claude-opus-5-5.

It accepts text and images, produces text, and supports tool use. The API lists a 1M-token context window and 128K maximum output tokens. See the model documentation.

Adaptive thinking is always enabled, with medium effort as the default. You can adjust the effort, but you can no longer switch to a separate non-thinking mode, bringing it in line with other reasoning models on the market. Not sure that's a good thing or bad thing, but it was always quite confusing that the Claude family of models had an effort setting that was separate from its reasoning capability.

Benchmarks

These are results reported in Anthropic's system card, Table 8.1.A. I've added comparisons to other similar models, all based on the vendors' reports.

The default evaluation uses adaptive thinking at max effort, averaged over five trials. Terminal-Bench 4.0 uses xhigh. These are higher effort settings than the API default. The launch results also include production safeguards, with fallback models handling some restricted tasks.

Opus 5.5 does not lead every comparison: the same table puts GPT-6 Astra ahead on Terminal-Bench-Science 0.1 and AutomationBench. Results from other developers use their published setups, so this is not a controlled comparison with an identical setup Coding Harness.

The system card also reports a useful caveat: despite improved resistance to prompt injection overall, Opus 5.5 is more susceptible to malicious instructions hidden inside text a user pastes into their own message. Better benchmark results do not remove the need to distinguish instructions from untrusted content.

Pricing and Cache Settings

Standard Claude API prices, checked on 23 September 2026, in USD per million tokens:

Input and output are 20% cheaper than Opus 5; cache reads are 60% cheaper. Reusing a large context can therefore make a bigger difference than the headline input price suggests. Pricing documentation.

Caching requires opting in through cache_control, either for automatic caching or explicit breakpoints. The minimum cacheable prefix is 512 tokens. The default lifetime is five minutes, refreshed when reused; a one-hour cache costs more to create. The example below makes one request without caching. Prompt caching documentation.

Real example

I used Opus 5.5 to turn this article into a 3D gameshow with an N64 look.

This uses the Anthropic Python SDK and my Obsidian Markdown Notebook plugin. A small helper saves the HTML and token usage, then renders the game. The model call is just a normal SDK call, with a one-sentence system prompt.

If you want to run it yourself, download the Notebook HTML helper, then install it in your notebook's Python environment:

pip install anthropic ./notesbylex_notebook_html-0.1.1-py3-none-any.whl

Set ANTHROPIC_API_KEY in the environment used to launch Obsidian. The example reads it from the notebook process's environment and passes it to the SDK, which sets the X-Api-Key header. Running the cell makes a paid request; the saved game plays without an API key or another model call.

import os
from anthropic import Anthropic
from notebook_html import read_article, render_anthropic_html

api_key = os.environ["ANTHROPIC_API_KEY"]
article = read_article("claude-opus-5-5.md", before="Real example")
with Anthropic(api_key=api_key).messages.stream(
    model="claude-opus-5-5",
    max_tokens=128_000,
    output_config={"effort": "high"},
    system="Return a complete, self-contained HTML document.",
    messages=[{"role": "user", "content": f"""
Turn this article into a 3D gameshow that looks like an N64 game.
Use around 8 questions that test understanding of the article, with a short
explanation after each answer.
Support phones and keyboards. Use inline CSS and JavaScript, no external assets,
network requests or browser storage, and no em dashes. Keep the code compact,
with one shared question renderer.

Article:
{article}
"""}],
) as stream:
    response = stream.get_final_message()
render_anthropic_html(response, "../_media/claude-opus-5-5-quiz-high.html")

This game was generated on 23 September 2026 using high effort. It cost about US$0.96, calculated from the saved token usage and standard API prices:

Of those output tokens, 36,150 were thinking tokens and 11,392 were the HTML response. Thinking is included in the output cost.

Play the quiz in its own page.

The example sets effort to high. To render the saved game again without calling the model, use load_html("../_media/claude-opus-5-5-quiz-high.html") from notebook_html.

Sources

• Claude Opus 5.5 announcement
• Claude Opus 5.5 system card
• Model specifications
• API pricing
• Prompt caching
• Steering thinking