Claude Opus 5.5 + Blender MCP

Anthropic released Claude Opus 5.5 on 22 September 2026. Anthropic says  it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” This page covers how to use it with Blender MCP, what it costs, and the API changes that matter if you drive Blender with your own code.

Claude Opus 5.5 at a glance

All values are from Anthropic’s models overview , migration guide , and launch page .

Does Claude Opus 5.5 work with Blender MCP?

Yes. Blender MCP does not depend on a particular model, and Anthropic lists tool use as supported on all current Claude models. The Blender addon and the MCP server offer the same tools (scene inspection, Python execution, object and material changes, and asset imports) to any client that supports the Model Context Protocol, and the client decides which model calls them. You do not need to change anything on the Blender side.

What a new model changes is how well it uses those tools: how accurately it reads the scene, how often its Python code runs on the first try, and how well it plans a scene that takes many steps. We have not published our own Blender test of Opus 5.5 yet, so run the quick test below on your own scenes.

How to use Claude Opus 5.5 in Blender

If you have not connected Claude to Blender yet, follow the Claude + Blender MCP setup guide first. It takes roughly ten minutes if you already have Blender installed.

API changes that affect Blender tool calls

If you use Claude Desktop or Claude Code, the client handles these for you. If you call the API yourself, Anthropic’s migration guide  lists four breaking changes from Opus 5. These are the ones a Blender MCP client is most likely to hit:

• Forced tool use returns a 400 error. tool_choice types any and tool are not supported. If your client forces a call such as get_scene_info before each turn, switch to auto, add "strict": true to the tool definitions, and say in the prompt when to use the tool.
• Thinking can’t be turned off. thinking: {"type": "disabled"} and manual budget_tokens both return a 400 error. Remove the thinking field and pick an effort  level instead. Pass thinking blocks back unchanged with each Blender tool result.
• Keep the conversation append-only. Thinking blocks are tied to the conversation that produced them. If your client rewrites the system prompt, the tool list, or earlier messages between Blender calls (for example to inject a fresh scene summary), replaying thinking blocks afterwards returns a 400 error by default for accounts created on or after 31 August 2026. Add new scene information as a new message instead.
• The default effort is lower. It is medium on Opus 5.5 and was high on Opus 5. If long scene builds get worse after you switch, set the effort level explicitly instead of relying on the default.
• Text between tool calls moved. The short notes the model writes between tool calls (“Now adding the camera…”) come back as thinking blocks, not text blocks. They are empty by default. If your interface shows progress while Blender works, it will go quiet. The migration guide explains the display setting that restores them.
• Computer use changed. If you drive the Blender interface with computer use instead of MCP, computer_20251124 returns a 400 error on the Claude API and Google Cloud. Use the computer_toolset_20260801 toolset instead. On Amazon Bedrock the old tool still works.

Claude Opus 5.5 vs Opus 5 vs Fable 5.1

Anthropic’s launch page  reports these benchmark results; prices are from the models overview . The benchmarks measure coding and computer use, not Blender, so treat them as a starting point.

Anthropic says Opus 5.5 “will cost 40% less than Opus 5 on typical workloads” at default settings. Anthropic attributes the saving to lower token prices and fewer tokens per task. What a Blender session actually costs depends on how many tokens it uses and how much of it is cached.

A quick test for Claude Opus 5.5 in Blender

Send the same prompts to Opus 5.5 and to the model you use now, in the same scene, and compare the results:

1. Scene reading. “List every object in the scene with its type, location, and material.” Check the answer against the Outliner.
2. An edit that takes many steps. “Add a ground plane, three cubes in a row with 0.5 m gaps, and a camera that frames all of them.” Count how many attempts it needs.
3. Recovery from an error. Ask for something that needs Python, such as “array this object along a curve.” See if the model fixes its code when Blender returns an error.

Save the .blend file before each run. Blender MCP can run arbitrary Python in your session. For a comparison of Blender AI tools in general, see the best Blender AI assistants.