Claude: Elevated errors across many models [resolved]

Original Article ↗ Hacker News Discussion ↗

Frequent outages and elevated error rates on Anthropic’s Claude models are frustrating users, especially those relying on long-running coding sessions in Claude Code. Commenters contrast Claude’s capabilities with persistent UI bugs, unreliable terminals (notably with tmux and Windows), and weaker uptime than OpenAI, and they question Anthropic’s claims that “coding is solved” when its own infrastructure and tooling remain fragile. Many report shifting more work to competitors like GPT/Codex, arguing that large-scale engineering and reliability are far from being automated away.

Service reliability & outage impact

  • Multiple users report Claude being unusable during the incident; some switch to other providers (e.g., GPT 5.5, Codex) citing better stability or speed.
  • Others say they rarely or never see issues, or feel reliability has improved in the last 1–2 months.
  • Some complain outages are now “every few weeks” and describe Anthropic as a “hot mess” since late 2025, though others disagree.
  • One user notes only Auto Mode’s safety classifier was erroring; switching to “Accept Edits” avoided failures.
  • Long-running agent sessions break on 500s; subagents erroring can cause the main agent to panic and reset the git state.

Claude Code quality, terminals, and UX

  • Several reports of Claude Code being slow, choppy, or garbling lines, especially in Windows Terminal and tmux; others say it’s “wonderful” and rock solid on their setups.
  • Specific bugs mentioned: massive scroll flashes, severe slowdowns in tmux over time, rendering issues that improve when vim mode is disabled.
  • Some compare other harnesses (Codex TUI, OpenCode, pi.dev) as faster, less buggy, or simpler; others argue Claude Code’s extra complexity reflects useful features (memory, tasks, sub-agents, worktrees) rather than pure bloat.
  • Complaints also extend to the web UI (opening settings spikes CPU/fans) and CLI flickering.

“Coding is solved” debate

  • Many criticize Anthropic’s public claims that coding/software engineering is “solved” or that “software engineer” as a role is fading.
  • Counterpoint: LLMs can write code, but complex engineering, large-scale distributed systems, and infrastructure reliability remain hard.
  • Some users see 10x gains for hobby projects but minimal gains on large, complex production systems, warning against unrealistic productivity expectations.

Model comparisons & usage

  • Mixed views on Claude vs GPT 5.5 vs Codex:
    • Some prefer Claude’s capability, humor, and lower BS confidence.
    • Others prefer GPT 5.5’s denser, more deterministic output and Codex’s long-task performance and auditability.
  • Claude is seen as stronger for rapid UI/frontend generation; Codex favored for large codebases and long tasks.
  • Outage risk is leading a few users to cancel Claude subs and rely solely on Codex or GPT.

Infrastructure, billing, and government offering

  • Users note Anthropic’s uptime lags OpenAI’s, and criticize janky products despite strong revenue and PMF.
  • One detailed comment describes confusing infrastructure/billing variants (Bedrock legacy/new, Claude for AWS, Marketplace, Anthropic’s own infra) and missing/buggy support for features like “Fast Mode” under certain billing routes.
  • Some speculate outages are due to rapid scaling, GPU scarcity, or “slop code” (referencing a prior code leak discussion), but root causes are unclear.
  • “Claude for Government” shows better uptime on the status page; users hypothesize separate, more predictable infrastructure.

Developer environments tangent

  • Long side discussion on Windows vs macOS/Linux as dev platforms.
  • Many report Windows is still common in large orgs, gaming, C++/graphics, CAD/CAE, and enterprises; others say they rarely see Windows among “serious” devs.
  • WSL2 is frequently mentioned as making Windows a viable dev environment.