Last week an indie developer burned through Claude Code's monthly token quota — and that tells us more than it might seem: the real cost of AI coding tools likely far exceeds the subscription number on the surface.
So he made a decision that looks radical but is actually pragmatic: invent a compressed programming language built specifically for large models.
What this is
This developer open-sourced TL (token-efficient language) on GitHub. The logic is direct: the curly braces, indentation, and long variable names in existing programming languages were all designed for human eyes — to large models, they're pure redundancy. TL strips them out, compressing code volume by 30-60%, then uses a VS Code plugin to restore the compressed code back into a human-readable version in real time — LLMs see the compressed version, humans see the normal version. Currently supports NodeJS only.
Industry view
Most on Reddit and Hacker News find the approach clever, and at least see it as proof that token consumption is a real pain point. But our editorial team is more interested in the dissent: this is using engineering to work around a problem the model itself should be solving. If large models find standard syntax too verbose, the model should be optimized — not humans forced to switch languages and write everything twice. The deeper risk: once this approach becomes popular, "code for humans" and "code for AI" will split into two systems, driving up long-term maintenance costs.
Impact on regular people
For enterprise IT: token consumption in AI coding tools is a hidden cost. When procuring, don't just stare at the subscription fee — estimate real call volume.
For working professionals: developers who heavily use AI to write code may burn through monthly quotas faster than expected. Start paying attention to usage management.
For consumer markets: no direct short-term impact, but this is a signal — AI service pricing models are still in violent flux. Don't treat today's prices as the new normal.