Elevated errors on Claude.ai, API, Claude Code

Original Article ↗ Hacker News Discussion ↗

Frequent errors and outages on Anthropic’s Claude.ai, API, and Claude Code are disrupting developer workflows and exposing how dependent many teams have become on these tools. Commenters highlight shrinking quotas, opaque pricing and capacity management, and reliability problems ranging from 500 errors to login failures, prompting some to keep backups like OpenAI’s Codex, GLM models, or local LLMs. The thread underscores a wider concern that frontier AI services are hitting infrastructure limits, with business models and technical architectures struggling to keep pace with surging demand.

Service reliability & 500 errors

  • Many users report frequent 500 errors on Claude.ai, API, and Claude Code, especially around 14:30 UTC / US West Coast morning.
  • Outages often affect everything, including auth/login, not just inference.
  • Some see patterns: peak-time overload, issues with caching, or new features (e.g., routines rollout) coinciding with instability.
  • Uptime estimates differ: some third‑party trackers show ~89–91% for Claude Code, while Anthropic’s official status shows ~98–99% over longer windows, leading to debate about how uptime is calculated and perceived.

Impact on workflows

  • Users relying on Claude Code for active development are significantly disrupted, especially during time‑sensitive work.
  • Several describe needing to “fall back to their brain,” textbooks, or traditional coding when outages hit.
  • People seek ways to export or recover interrupted sessions; built‑in export sometimes fails, prompting DIY solutions (JSONL parsing, SQLite logging, custom tools/skills).

Alternatives & backups

  • Codex (OpenAI + desktop app) is the most cited alternative; generally praised for reliability, execution quality, and more generous or clearer quotas, though some find it less “refined” or weaker at planning.
  • Other options mentioned: Gemini/Vertex AI, GitHub Copilot, Qwen, GLM via z.ai, Fireworks-hosted models (including Kimi), OpenRouter, Grok, Turnstone with local/open models, llama.cpp.
  • Several users now tool repos to support both Claude Code and Codex (or others) and switch on outage.

Pricing, quotas & “shrinkflation”

  • Reports of rapidly hitting Claude quotas, especially after changes to reasoning defaults and peak‑hour limits.
  • Complaints that trivial tasks now burn through large chunks of allowance; some upgraded plans just to keep working.
  • Comparisons to OpenAI’s and others’ pricing; some perceive token “shrinkflation” and pressure toward higher tiers (e.g., $100 / $200 plans).

Product & support criticisms

  • Users praise Claude Code’s core coding ability but criticize:
    • Poor or slow support responses.
    • Confusing account/organization separation.
    • Broken payments and login/token flows (including CLI/CI issues).
    • Buggy desktop app that “forgets” it’s the desktop environment.
    • Weak or opaque MCP integration and debugging.
  • Some feel early adopters are being squeezed as usage grows.

Status page & transparency

  • Mixed views: some say Anthropic’s status page is relatively honest vs big clouds; others note lagging updates and “all systems operational” while 500s occur.
  • A few suggest more openness about capacity constraints, GPU shortages, and policy changes.

Suggested mitigations & broader reflections

  • Ideas: queuing instead of hard errors, surge pricing during peak, automatic fallback to partner/cheaper models, disabling heavy features (e.g., routines) at peak.
  • Meta‑discussion that developers have become highly dependent on AI tools in under two years, making outages feel like “GitHub going down” used to.