Claude: Elevated errors across many models [resolved]
TL;DR Highlight
Claude Sonnet and Opus had roughly 10% error rates during an incident lasting about two hours.
Who Should Read
Developers running Claude API or Claude Code in production or long-running agent jobs who want to handle outages better.
Core Mechanics
- Two phases of the incident
- From 10:23 PT (17:23 UTC) to 11:00 PT (18:00 UTC) on 2026-06-16, all Sonnet and Opus models were affected.
- Error rate reached around 10% in this window.
- Then only Opus 4.8
- From 11:00 PT (18:00 UTC) to 12:20 PT (19:20 UTC), Opus 4.8 averaged a 10% error rate.
- An interim update said errors persisted on Opus 4.8 and Haiku 4.5.
- Wide impact
- claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork were all affected.
- Response timeline
- Investigation began at 17:29 UTC and a first fix landed at 18:00 UTC.
- A second fix landed at 19:22 UTC and the incident was marked resolved at 19:32 UTC.
- No root cause given
- The status page includes no RCA or prevention measures.
Community Reactions
- Experience of a main agent panicking on subagent 500s
- One commenter said a long session spawned subagents, and when one hit a 500 error, the main agent did a git checkout to restore a known-good state.
- Experience of errors only in Auto Mode
- One commenter saw errors only from Auto Mode's safety classifier, while Accept Edits mode ran the same bash operations fine.
- Complaint about long sessions being broken
- Commenters said the errors bust long-running sessions and make them unusable.
- Opinions about switching to Codex
- Several commenters said they moved to Codex with GPT 5.5 or went back to gpt-5.5 during the outage.
How to Apply
- Add retries with backoff to API calls
- Today, wrap your Claude calls with exponential backoff and a max retry count for 5xx responses.
- At around a 10% error rate, retries absorb most failures.
- Block destructive commands on agent errors
- Instruct subagents not to run git checkout or reset when they get an API error.
- Make them distinguish API errors from task failures, retry, then report.
- Build a model fallback path
- Prepare a config to switch models, since only one model (e.g. Opus 4.8) may be failing.
- In phase two of this incident only Opus 4.8 was affected.
- Checkpoint long-running jobs
- Save progress to a file at each step so a session cut by errors can resume.
Try It
# Paste into Claude Code or a Claude/ChatGPT chat: agent error-handling rules
When an API error (5xx, timeout, overloaded) occurs during work, follow these rules.
1. Do not run revert commands such as git checkout, git reset, or file deletion.
2. Retry the same call up to 3 times at 5s, 15s, and 45s intervals.
3. If it still fails, stop and record current progress and the failed step in {{progress file}}.
4. Report to the user which subtask failed with which error, then wait for instructions.
Current task: {{task description}}Terminology
Related Papers
AI isn't replacing SREs, it's deskilling them
Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.
Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor
Building an AI-native engineering org
Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%
We ran AI agent swarms in Minecraft and Mindustry
We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.
The Agent SDK
Cursor released Agent SDK, letting you define agents in code and deploy them with one command.
Here be monsters: What do you still need to know when agents do the work?
Owning agent-written code depends on fundamentals, breadth, and diving deep just in time.
AI promised less work. I've never worked more
Running agents in parallel brought my GitHub activity back to full-time engineer levels, and I'm exhausted.
Show HN: Corgi, a helpful {agent, terminal} multiplexer
A tmux-like terminal tool that shows the status of all your AI agents at a glance