Ultracode and Ultrathink for OpenCode. Fusion beats Fable

Anthropic's self-inflicted sanctions triggered the US authorities to act based on the Wassenaar accords. Their Fable model was declared a national security threat (probably based on interesting procedures). This gives even more market to open-weight models and empowers international companies to use alternative models, including Chinese open-weight models.

Good news: the EU isn't limited. Bad news: the EU cannot concentrate capital to produce competitive models locally. But that isn't a new problem. It's related to the ways of European fiscal policies.

What is Fable? And what is Ultracode?

Officially Fable is a version of Anthropic's "Mythos-class" flagship / frontier / expensive model. I however believe Fable to be a fusion of Opus models. The performance is too similar, and this cannot be a coincidence.

The actual innovation in Claude Code is /ultracode. Not Fable. Fable is just another token predictor. Ultracode is a novel way to solve coding problems.

And this is now available in general:

Ultracode

Ultracode is a multi-agent way to get better results for intensive coding tasks. I have implemented /ultracode mode for OpenCode. You can use it with any compatible model (literally 100s).

I have tested it with

MiMo V2.5 Pro
GPT 5.5 xhigh
DeepSeek V4
Qwen 3.7 Max

For security reasons I cannot guarantee Mistral support to my US friends. But I will reach out to trusted EU industry partners. Together we will fight the slop. And hackers. Of course. Slop hackers. Using expensive AI.

To be fair: the ultracode plugin relies on OpenCode's agentic framework. It works best with dynamic context pruning and Skills (which contain your coding guidelines for the project).

Ultrathink

Ultrathink is the same, but for research and development tasks. That is something I haven't seen on Anthropic's capability list. I added it. You can use it with the Skills I have published:

for some of the intelligence analysis works it's best to use grounded models (!)

Generally I think adding multi-agentic reasoning to these Skills is a good addition.

Fusion

Fusion is another multi-agent approach. Ultrathink / Ultracode is about dispatching agents intelligently for the tasks at hand.

The fusion variants exhaust the probabilistic spectrum of the model answers deeper. This way for example Qwen 3.7 Max can be used in addition to GPT 5.5 xhigh. You can also fuse answers of the same model.

Or you could use open-weight models and beat Fable directly. Try it out.

Summary

initial release of OpenCode plugins for Ultracode, Ultrathink and Fusion
CI / CD for opencode plugin, based on Dagger

With UltraCode and Fusion you can get the AI experience you deserve. More efficient than Anthropic models, cheaper and faster. Less marketing, more substance.

My favourite open-weight combo is DeepSeek V4 Pro + Xiaomi MiMo 2.5 Pro. Alibaba's Qwen 3.7 Max behaves oddly. Tencent's Hy3 looks promising as a future arbiter model to replace OpenAI GPT 5.5.

The most important fusion is with me, and not between models.

AI is an opportunity, not a gift

US AI is slowly drifting into a crisis. The regulations (I thought that's a EU thing), empty marketing promises and cost explosion won't vanish. 2027 may be the year of Chinese AI rather than of successful US IPOs.

Demo of Ultrathink Fusion in OpenCode

Note: you can also use OpenCode Desktop (on Windows).

/ultrathink-fusion --panel ultracode-fusion-panel-a --panel ultracode-fusion-panel-b --decider selected-model "What is the difference between idiosyncratic and macro-economic risk for a factor invest portfolio"