Companies Are Making Claude, Codex Talk Like Cavemen to Stop AIs Soaring Costs
To reduce soaring AI inference costs, some companies are imposing strict token limits on models like Claude and Codex, forcing them to communicate in extremely truncated, "caveman-like" language. This technique cuts operational expenses sharply but risks degrading output quality and limiting the AI's ability to reason through complex tasks.
Background
- Anthropic's Claude and OpenAI's Codex are among the most capable (and computationally expensive) LLMs on the market. "Talking like cavemen" refers to a technique called *token optimization* or "prompt compression": stripping articles, conjunctions, punctuation, and polite phrasing from prompts so the AI generates fewer tokens — each of which costs the company money per API call.
- For big customers (e.g., coding assistants, customer support platforms), these savings add up fast, but the trade-off is that the model's reasoning quality often degrades: crucial nuance, safety guardrails, and formatting reliability can break when prompts become garbled.
- The phenomenon highlights a tension in the current AI industry: frontier models are powerful but so expensive to run at scale that users feel forced to degrade the experience to keep budgets manageable.
- This is a 2025-era story, not an early experiment — it reflects real, widespread deployment pressure now that AI is embedded in production software, not just demo products.