Anthropic has drawn a line around video games in Claude Code. According to Tom’s Hardware, an Anthropic employee confirmed that Claude Code now has orders not to reproduce game IP, and the company says it is “iterating to ensure […] the right balance.” At the same time, users are reporting that reverse engineering tasks are being refused. The blocking appears limited to one combination: Claude Code running on Opus 5.5.
I don’t have a problem with the first half of that. I have a real problem with the second.
The game-cloning rule is defensible
Agentic coding tools have become good enough that “build me a copy of this game” is no longer a joke prompt. An agent that can scaffold a project, write the gameplay loop, and iterate on its own output can get surprisingly close to a commercial title’s mechanics and presentation. If you’re Anthropic, you don’t want your flagship coding product to become the easiest way to pump knockoffs of someone else’s work into storefronts.
So a guardrail against reproducing existing game IP makes sense. Characters, names, art direction and recognizable level designs belong to the people who made them. Telling an agent not to recreate those wholesale is a reasonable call, and frankly it’s surprising this didn’t arrive sooner.
Genres are a different matter. A platformer is a genre. Nobody owns tile-matching or top-down shooting. Whether the new rule can tell “make a game like X” apart from “make X” is the thing developers should be testing right now, because the sources don’t say where Anthropic has put that boundary.
Reverse engineering is the wrong casualty
The user reports about reverse engineering are where this goes off the rails. Reverse engineering is ordinary engineering work. People do it to understand file formats, get old software running on new hardware, build mods, write compatibility layers, debug binaries they lost the source for, and audit code for security holes. Much of that has nothing to do with games, and the parts that do (modding, preservation, interoperability) are some of the healthiest corners of PC gaming.
A coding agent that flinches at disassembly or protocol analysis is a worse coding agent. If the anti-cloning rule is spilling into this territory, that is a calibration failure, and Anthropic’s own “right balance” language reads like an admission that the current setting is off.
The company should fix it quickly and say so publicly. Developers choose tools based on what they’ll refuse, and quiet refusals erode trust fast.
Why only one combination?
The narrowness is the strangest detail. Reports point to Claude Code with Opus 5.5 specifically, which suggests this is a product-level policy layered onto one pairing rather than a change to Anthropic’s models across the board. That would also make it easier to tune, which is good news if Anthropic is serious about iterating.
It also means the restriction is inconsistent by design. A developer blocked in Claude Code on Opus 5.5 may get a different answer by switching models or surfaces. That patchiness frustrates honest users without doing much to stop determined cloners.
Anthropic hasn’t published a list of what counts as game IP for the purposes of this rule, or which reverse engineering tasks are meant to be off-limits versus caught by accident. Until it does, nobody using Claude Code on Opus 5.5 knows whether asking it to parse an old save-file format will work.