Elevated errors for multiple models – Resolved
Elevated error rates on Anthropic’s Claude models triggered complaints from developers who lost work mid-session and highlighted how dependent many workflows have become on hosted LLM APIs. Commenters debated accountability for frequent outages, speculation that instability precedes major model releases, and increasingly aggressive safety filters that interrupt legitimate technical and scientific use cases. Many described pivoting to rival models or building tooling to switch harnesses and maintain local fallbacks, underscoring reliability and vendor lock-in as growing concerns in the AI tooling ecosystem.
Outage scope and possible causes
- Users report elevated errors on multiple Claude models; some also see issues with Grok, suggesting shared infrastructure problems (e.g., “Colossus”, AWS us-east-1).
- Atlassian Statuspage is used; some note subscription widgets briefly malfunctioned.
- Several wonder if AWS issues (linked health status page) contributed, but this remains unclear.
Reliability, SLAs, and economic impact
- Frustration about lost work, burned tokens, and time, with no refunds or acknowledgments.
- Some say such outages would be unacceptable for core services like search; others note humans (engineers) also have imperfect availability.
- People ask about formal SLAs and post‑mortems; none are discussed in the thread.
- Commenters argue outages underline the need for local fallbacks, smaller local models, and resilient workflows.
Tooling, harnesses, and multi-model workflows
- Many describe seamlessly switching between harnesses (Claude Code, Codex, Cursor, Pi, others) and models (Claude, OpenAI, DeepSeek, etc.) when one fails.
- Techniques include storing state in local files or git, custom LLM routers, “handoff” prompts/skills, and session-resume features.
- Some think dedicated products for cross-harness switching are overkill since a few prompts or scripts can handle it.
Model quality, pricing, and competition
- Some users shift toward OpenAI (e.g., Astra) due to perceived speed, price, or access, though limits on usage are noted.
- Others describe multi-model “adversarial review” workflows where one model audits another’s design output.
Safety safeguards and access limits
- Multiple users report sudden increases in “safeguard” blocks, especially for biology, cybersecurity, or even benign terms (e.g., “microbes”).
- People speculate safeguards were tightened after a possible incident, but this is unclear.
Speculation on new models and rapid release cadence
- Several link outages to rumored upcoming releases (e.g., Opus 5.5, Fable 5.2), citing behavior of API routes (403 vs 404) as informal signals.
- The rapid stream of new models is compared to 1980s/2000s hardware “wars”; some find it exciting, others exhausting.
Meta: HN coverage and corporate power
- Some complain that every AI outage hitting the front page crowds out “things people are building.”
- Others respond that upvotes reflect interest, like Apple events or local SF news.
- A long subthread criticizes concentration of power and wealth in tech executives and argues for holding individuals, not just corporations, accountable.