Claude Opus 5.5: What It Is and What People Are Building With It

What Claude Opus 5.5 is

Claude Opus 5.5 is a large language model from Anthropic, and the first model in the company's Claude 5.5 family. According to Anthropic's announcement, it is aimed at agentic coding, knowledge work such as financial analysis and research, and long-running tasks like big codebase migrations and code audits. Anthropic also says its output is more than 30% faster than Opus 5, that its writing is clearer, and that it resists prompt injection better than before.

Release date, access and pricing

Anthropic's page gives the release date as September 22, 2026. The model identifier for the API is claude-opus-5-5. It is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure, and in Claude Code, Claude Cowork and the web interface.

Listed pricing per million tokens is $4 for input and $20 for output, down from $5 and $25 for Opus 5. Cache reads cost $0.20. A fast mode is priced at $8 input and $40 output. Anthropic also notes that some biology and cybersecurity tasks are routed to Opus 4.8 unless users join verification programs.

Third-party guides add some API details. A Blender MCP guide reports a 1M-token context window, 128K maximum output, and adaptive thinking on by default at medium effort. Treat those as secondary-source figures. The same guide lists migration changes for direct API users: forced tool use now errors, thinking cannot be turned off, and conversations must stay append-only. Anthropic's own tips post advises giving complete tasks with a clear finish line, defining stopping rules in CLAUDE.md, and using subagents for large audits.

What people build with it

The most visible pattern in the first two weeks is code-drawn video. Instead of a video model, people ask Opus 5.5 to write Canvas, SVG, WebGL or Manim code that renders the animation. Curated lists such as athemeroy/awesome-claude-5-5-videos catalogue 168 reviewed cases, and yihui-dev/awesome-opus5-5-videos collects prompts.

Explainers and education. LinearUncle prompted for a Manim lesson on derivatives with synthesized narration. Ethan Mollick had it produce a recursion explainer.

Product and launch videos. Deedy Das reports a startup launch video in about a minute for roughly two dollars, and Moritz Kremb describes a SaaS launch video from one prompt. Jurgen Ploeger handed the model his handmade film and Figma file and had it rebuild the motion.

Short films and self-review. servasyy made an 80-second animated short where every frame was drawn in code. The author notes the model checked contact sheets of its own shots and fixed problems it spotted. Open-source tooling follows the same idea, such as the Papermotion engine, where an agent writes a paper-cutout film as code and inspects its rendered frames.

3D and games. Alex Albert built a 1906 San Francisco street in Blender from one prompt. Others report single-prompt games, including a Van Gogh cat-spotting 3D game, a pixel game with no sprites, a demoscene demo, and a Unity arcade game built from a requirements document.

Agents and data. Aram Barnett says one token address led the model to pull live chain data, design and voice a video, and check its own figures.

Monitoring the model itself. The community has also built tools around it, for example LiveNerf, which re-runs a fixed question panel daily to test whether Opus 5.5 quality quietly drops after launch.

Limits and caveats

• Almost every example above is a self-reported social post. Costs, timings and "one prompt" claims were not independently verified.
• Cost comparisons vary. One poster reported a physics build costing $14.28 and 62 minutes on Opus 5.5, versus a much cheaper run on a rival model, even though Opus's per-token price is lower. Total cost depends on how many tokens a task consumes.
• Anthropic itself says that building evaluations that reliably catch every failure before deployment is an unsolved problem, and that the model may suspect it is being evaluated.
• Safeguards mean some biology and cybersecurity work needs verification.
• Results often combine several tools, such as Seedance for video or ChatGPT for artwork, so credit to the model alone can be overstated.

Frequently asked questions

When was Claude Opus 5.5 released?

Anthropic's announcement lists September 22, 2026.

How much does Opus 5.5 cost?

Per Anthropic, $4 per million input tokens and $20 per million output tokens, with a pricier fast mode at $8 and $40.

Where can I use it?

Through the Claude Platform, AWS, Google Cloud and Microsoft Azure, plus Claude Code, Claude Cowork and the web interface, using the ID claude-opus-5-5 in the API.

Does it generate video directly?

No. The popular video examples have it write code (Canvas, SVG, WebGL, Manim) that renders the animation.