Which language is best for AI code generation? (revelry.co)

Tencent’s AI R&D team released AutoCodeBench — an automated pipeline that generates and verifies 3,920 high-difficulty coding problems across 20 programming languages in sandboxed environments — and used it to benchmark modern LLMs. The headline finding: top models such as Claude Opus 4 and GPT‑4.1 average around ~50% Pass@1 overall, but Elixir stands out, exceeding ~80% Pass@1 in both reasoning and non‑reasoning modes. Because AutoCodeBench includes “low‑resource” languages and auto‑validated inputs, its results suggest LLM codegen performance depends heavily on language traits and training data, not just on Python/JavaScript popularity. Technically, the authors argue Elixir’s strengths come from functional programming features — immutability, small, composable standard libraries and consistent syntax — which reduce the context and hidden state an LLM must track. Elixir’s “Goldilocks” age and higher‑quality public code may also yield cleaner training signals. Caveats remain: a separate Stanford study found AI can hurt productivity in niche languages as task complexity rises, and model training/bias (RL fine‑tuning, dataset composition) matters a lot. Practical implications: language choice should factor in LLM friendliness, codebase manageability and tooling (prompting, opinionated formatters, runtime‑aware agents like Tidewave) rather than defaulting to Python/JS.