LLM vibe coders debate model power, token economics, and Bun's alleged deception

Users on the technology board gathered to discuss LLM-driven “vibe coding” workflows, autonomous agents, and the tools that promise to let non-programmers build software. The conversation surfaced recurring complaints about pricing, model capability gaps, and a sustained attack on Bun, the JavaScript runtime.

On Bun, one respondent alleged a sustained deception. “Massive scam,” the user wrote, claiming founder Jarred Summer’s performance benchmarks relied entirely on uWebSockets, a C++ library that actually runs faster in Node.js than in Bun. The commenter cited a commit from the uWebSockets team itself calling out the discrepancy, along with GitHub issues where the Hyper-Express creator allegedly demonstrated that realistic network-based benchmarks, rather than local testing, showed Bun’s claimed advantages evaporating. The user concluded: “The tranny of course didn’t respond or ever even fucking address it.”

Users also griped about model limitations. One commenter claimed “Opus got lobotomized somehow” and alleged that Claude Code remains “unusable,” while another stated that ChatGPT’s planning mode “is just a summary of the conversation so far, lacks a shitton of context and is unusable standalone.”

Token pricing emerged as a flashpoint. Users reportedly complained that Claude’s subscription tier “tokens are like half of what you get with OpenAI” and that neither the $20 nor $100 ChatGPT plans hit the sweet spot. One user claimed the “$20 plan is too few tokens… the $100 plan is too many tokens,” and expressed frustration that Anthropic’s pricing forced them to juggle multiple subscriptions alongside free Gemini.

On tooling, commenters disputed whether the underlying AI model or the agentic interface mattered more. One user argued “the tool doesn’t matter. Whatever model you use, it will perform more or less the same.” Another claimed Opus handles “lesser known software” better, while GPT-based agents “get confused running my programs, even misquotes shell commands from time to time.”

A respondent described using multiple models in parallel: ChatGPT for thinking and planning, Codex for execution, and Gemma 4 for verification, though they acknowledged the third tool likely adds no measurable value. Another user reported building a custom agent atop an open-source platform using Codex, describing it as functional but requiring ongoing token management and iteration.

One commenter reported a novel use case: a vibe-coding step in a job interview process where candidates are tasked with fixing a production bug using Claude. They sought advice on structuring prompts for maximum success.