Claude struggles to cope with ChatGPT exodus

Original Article ↗ Hacker News Discussion ↗

Switching and usage patterns

  • Many report moving between ChatGPT, Claude, and Gemini with almost no friction; code changes to swap APIs are minimal.
  • Several now use Claude as primary, others moved to Gemini or still prefer OpenAI; many keep accounts on multiple services.
  • Some note the current spike for Claude may be mostly free users, with unclear revenue upside.

Model quality & UX comparisons

  • Claude is praised as an excellent “collaborator”: asks clarifying questions, reasons about user intent, and feels more conversational. Criticisms: brittle limits, occasional “meltdown” behavior, bugs in desktop app/state machine, and need for close supervision on larger tasks.
  • OpenAI’s Codex is seen as strong, literal, and good for long, well-defined jobs. It’s described as “boring but reliable,” with fewer dramatics but sometimes weaker collaboration.
  • Opinions on GPT‑5.4 codex diverge: some find it surprisingly strong and test-focused; others call it poor on out-of-distribution tasks (e.g., nonstandard Bazel rules).
  • Gemini gets mixed reviews: good inside Google’s ecosystem and for code review according to some; others call it weak on real-world/complex work unless carefully configured (e.g., forcing Pro instead of router). Rate limits are a recurring complaint.
  • Other models: Grok praised for speed and goal-focus but shallow reasoning; Chinese models (DeepSeek/Kimi) described as less polished but more robust on very weird/novel problems.

Ethics, surveillance, and Pentagon deals

  • Strong debate over OpenAI’s government contract language: especially that protections are framed around “U.S. persons,” leaving non‑US users feeling explicitly unprotected.
  • Some see Anthropic’s “red lines” as meaningful (people were reportedly fired over them); others call them PR with limited substance and note Anthropic’s own defense work history.
  • Several argue neither major lab is clearly “good”; concern centers on surveillance, autonomous weapons, and perceived gaslighting or weasel words.
  • Others are fatalistic: military AI use is seen as inevitable, and consumer boycotts as largely ineffective.

Commoditization, pricing, and moats

  • Many treat LLMs as interchangeable commodities; vendor choice is driven by price, rate limits, and immediate task performance more than loyalty.
  • Some predict long‑term competition will focus on pricing and compute capacity rather than raw model IQ.
  • Proposed moats: infrastructure reliability, velocity of datacenter build‑out, integrated tooling/agents/GUI, and personalized “memories” across sessions.
  • Counterarguments: user profiles can be exported or quickly relearned; personalization and history are not yet deep, and the market resembles undifferentiated web hosting.

Reliability and limits

  • Anthropic is criticized for unstable limits and frequent 504s on Opus; some stick to cheaper tiers to avoid hitting caps.
  • Others note Claude Code subscription restrictions (e.g., using it via third‑party tools) as a competitive disadvantage.
  • OpenAI/Codex are perceived as somewhat more stable and generous in usage, though ethical concerns are pushing some users away despite better performance.