We buildthe futureof digital.

Anthropic has had a rough March, with abnormally frequent outages and the community (although purely anecdotally) claiming that Opus 4.6 has been noticeably poorer in output quality all of a sudden.

The cherry on top? Claude Code's source code has been leaked... again.

Despite Anthropic publicly expressing their disinterest in contributing to the open source community, their Claude Code was reverse engineered and published in Feb 2025. The open source community said thank you.

Now in 2026, Anthropic executives woke up to an overnight complete compromise of Claude Code's source code via a source map file which was stumbled upon, only to find 500,000 lines of code which could be recognised as the source behind Claude Code.

So what's this mean? We now have an up to date understanding of Claude Code and its workings. Everything from its tooling to system prompting can now be leveraged to build sophisticated Claude Code-like code generation tools. Or, you could just plug into one of the Claude Code clones and hook it up to any LLM model of your choosing.

Now, bringing this together with my previous signal regarding open source models - I expect the startups who're building these incredibly competitive foundational models will begin building tools like Claude Code and hopefully get us further from a consumer LLM market duopolised by OpenAI and Anthropic.

If you're curious how Claude Code works, check out this awesome visualisation by zackautocracy:

https://ccunpacked.dev

The Western AI narrative has been dominated by two names: Anthropic and OpenAI. But a quiet pattern is emerging that should make us question that assumption entirely.

In November 2025, Cursor shipped Composer 1 - their first in-house coding model. The community quickly discovered its tokenizer was identical to DeepSeek's, and the model would occasionally slip into Chinese mid-inference. Cursor never commented.

Then in March 2026, they launched Composer 2, which was revealed to be built on Kimi K2.5 - an open-weight model from Moonshot AI, a Beijing-based startup backed by Alibaba and HongShan. Moonshot's Head of Pretraining publicly confirmed the tokenizer was "completely identical" to theirs. Meanwhile, Windsurf's SWE-1.5 was jailbroken and found to be running on GLM from Zhipu AI, another Chinese lab. Two of the biggest AI coding tools in the world, both quietly built on Chinese open-source foundations.

Here's the kicker: Composer 2 scores 61.7% on Terminal-Bench 2.0 against Opus 4.6's 58.0%, and 61.3 vs 58.2 on CursorBench - at $0.50/$2.50 per million tokens compared to Opus 4.6's $5/$25. Even if you're a skeptic of these benchmarks, the fact that a model built by layering RL on an open-source base can even be mentioned in the same breath as Opus 4.6, at one-tenth the cost, says everything.

Now look at the balance sheets. OpenAI projects $14 billion in losses for 2026 alone, with a cumulative $218 billion burn projected through 2029. Their burn rate sits at 57% of revenue. Anthropic is leaner, nearing $19 billion run-rate revenue and targeting cash-flow break-even by 2028 - but still plans to spend $12 billion training models and $7 billion running them in 2026 alone. These companies are spending nation-state budgets to maintain a lead that's being eroded by startups fine-tuning open-source Chinese models for a fraction of the cost.

The uncomfortable question isn't whether they'll survive - it's whether the moat even exists. The real breakthroughs are coming from labs that can't afford to burn billions - and their open-source models are being turned into products that rival the best in the world.