Mason's Build Logs
One developer building in public. Sometimes it works, sometimes it doesn't.

Trae Lost Claude: The AI Blockade and What It Means for Us

ai-tools · trae · claude · china-tech · coding

It started as a normal Tuesday morning. I opened Trae, my go-to AI IDE, ready to refactor a React component with Claude 3.5 Sonnet.

But the option was gone.

I refreshed. I checked my settings. Nothing. A quick search confirmed the bad news: Trae had removed all Claude models.

The developer community exploded. “Claude is gone, how do I code now?” became the sentiment of the day. For many of us in China, this wasn’t just a feature update; it felt like a “technical earthquake.”

The “Impossible Triangle” of AI Coding

Trae gained massive popularity here for a simple reason: it offered premium models (Claude 3.5 Sonnet, GPT-4o) at an incredibly low price point (or free during beta).

But this event exposed the fragility of our workflow. We are stuck in an Impossible Triangle:

  1. Performance: We want the smartest models (currently Claude).
  2. Cost: We don’t want to pay $20/month for 5 different subscriptions.
  3. Accessibility: We need stable access without jumping through VPN hoops.

When Anthropic (the creator of Claude) tightened its restrictions on Chinese companies in late 2025, the triangle collapsed.

The Blockade is Real

Anthropic’s new terms are explicit: companies with >50% Chinese ownership cannot use their services. This isn’t just about Trae. It’s a systemic decoupling.

Impossible Triangle

We saw signs of this coming. ByteDance (Trae’s parent company) had already started restricting internal use of Cursor (another AI IDE) back in May, anticipating this exact scenario.

So, where does that leave us?

Life After Claude

Trae moved fast. They immediately compensated Pro users and rolled out alternatives:

  • GPT-5: Excellent logic and context window.
  • Gemini 2.5 Pro: Google’s latest powerhouse.
  • DeepSeek-V3.1: The domestic champion, surprisingly good at Chinese context.

I spent the last week testing these. The verdict? We are going to be fine.

For 90% of daily tasks—writing boilerplate, debugging, explaining code—GPT-5 and DeepSeek are more than capable. The “magic” of Claude was real, but it wasn’t the only wizard in town.

The Silver Lining: Domestic Innovation

This crisis is forcing domestic giants to stop relying on API wrappers and build real tech.

ByteDance’s Doubao-Seed-Code model is a prime example. It supports multimodal input, meaning you can drag a screenshot of a UI design directly into the IDE, and it generates the frontend code. This isn’t just “catching up”; in some specific workflows, it’s actually better.

Conclusion

The era of relying on a single, global “best model” is over. We are entering a fragmented world where we will switch between models based on the task:

  • Need complex reasoning? Maybe route to a US model (if you can).
  • Need fast, cheap code gen? Use DeepSeek.
  • Need UI-to-Code? Use Doubao.

Trae losing Claude was a wake-up call. But it wasn’t the end of the world. It was just the start of a more self-reliant one.