Opus 5.5 vs Fable 5.1 vs GPT-6 Astra Code Minecraft Plugin (Advanced Test)
Matej (kangarko) · 2026-09-24 · review · 25,718 views
What's in the video
Description written by Gemini, which watched and listened to the whole video.
Summary
Matej (kangarko) from MineAcademy benchmarks three frontier AI coding models—Anthropic’s Claude Opus 5.5, Claude Fable 5.1, and OpenAI’s GPT-6 Astra—on developing a full Spigot Minecraft plugin from scratch. The models are tasked with creating a feature-complete "Meteor Strike" plugin with GUI menus, animations, physics, world rollback, and 11-year backwards compatibility spanning Minecraft 1.8.8 (2015) to modern Minecraft 26.3. Matej inspects the generated Java code in Eclipse IDE and live-tests each plugin on both modern and legacy Minecraft servers.
What is shown
- [00:20] The test specification: building a "Meteor Strike" plugin with custom GUIs, particle/sound animations, craters, block rollback, and cross-version compatibility between Minecraft 1.8.8 and 26.3.
- [01:23] Presenter's configuration files on GitHub (
github.com/kangarko/ai-files), including customCLAUDE.md, system prompt guidelines, skills, and Mineflayer bot integration for autonomous server testing. - [01:51] Brainstorming and drafting the detailed multi-phase storyboard prompt in Claude.
- [04:01] Running Claude Code in the terminal to autonomously generate and self-test the Opus 5.5 implementation.
- [06:01] Code review of the Opus 5.5 build in Eclipse IDE, highlighting modular class design (
CompSound,CompMaterial,Remainreflection bridge for legacy NMS handling). - [09:02] In-game test of Opus 5.5 on Minecraft 26.3: polished animated GUI, meteor target selection, impact countdown, crater explosion, and block rollback.
- [10:45] In-game test of Opus 5.5 on Minecraft 1.8.8: execution works, but reveals a client-side block desynchronization bug during terrain restoration.
- [12:12] Code review of Claude Fable 5.1: monolithic class design (
Strike.javaspanning over 500 extra lines), cleaner reflection handling. - [14:18] In-game test of Fable 5.1 on 26.3 and [16:00] on 1.8.8: clunky GUI layout, but smooth night-cycle transitions, particle effects, and no block desync on 1.8.8.
- [17:16] Code review of GPT-6 Astra: generated in a single crammed file (
BukkitVisuals.java) with silent exception swallowing and poor separation of concerns. - [19:54] In-game test of GPT-6 Astra on 26.3 and [22:05] on 1.8.8: tacky menu styling, unneeded target button, functional impact sequence, but an unexpected teleport bug on 1.8.8.
- [22:40] Final rankings: Opus 5.5 in 1st place (superior code taste and GUI polish despite legacy desync), Fable 5.1 in 2nd place, and GPT-6 Astra in 3rd place.
Claims & numbers
- The challenge tests cross-compatibility across 11 years of Minecraft updates (version 1.8.8 released in 2015 up to modern version 26.3).
- Models tested: Claude Opus 5.5 (max effort), Claude Fable 5.1 (max effort), and GPT-6 Astra (ultra effort).
- Matej states that GPT-6 Astra took over an hour to complete the task, making it the slowest model tested [17:23].
- Fable 5.1’s
Strike.javais about 500 lines larger than Opus 5.5's split classes [12:44]. - Matej rates Opus 5.5's GUI animation a 10 out of 10 [09:17], while rating GPT-6 Astra's code architecture a 4 out of 10 [13:35].
- MineAcademy has been running developer courses for 7 to 8 years [23:43].
Notable quotes
- [01:02] "Are you ready? Let's burn some tokens."
- [10:23] "This is near perfection and this is one-shotted all the code."
- [17:23] "Took more than an hour for GPT-6 Astra, it's actually the slowest one."
Assessment
This is a genuine, hands-on independent review and coding benchmark comparing three LLMs on a demanding real-world software engineering task. The code inspection and server runtime demonstrations are shown live on screen without visible cuts during gameplay testing, providing an authentic look at each model's code quality and execution flaws.
Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.