Anthropic Claude API — Complete Error Reference
Every documented error across the Anthropic Claude API, tested and verified by our engineering team. Rate limits, context window overflow, tool use validation, extended thinking, prompt caching, and MCP integration — with working code for every fix.
About the Anthropic Claude API
The Anthropic Claude API powers the Claude family of large language models — Opus, Sonnet, and Haiku — accessible via api.anthropic.com, Amazon Bedrock, and Google Vertex AI. Since its expansion in 2023, Claude has become one of the three most-used commercial LLM APIs alongside OpenAI and Google Gemini, particularly favored by developers building agentic workflows, long-context applications, and safety-critical products.
Because Claude's API surface is large — spanning traditional messages, streaming, tool use, extended thinking, prompt caching, MCP integration, and multiple deployment targets — the error surface is equally large. Most developers hit the same 20 or so errors repeatedly. This hub organizes all 254 documented errors so you can find your fix in under a minute.
Start with the official Anthropic quickstart. Come back here when things break.
Popular Errors Right Now
These are the eight most-searched Claude errors this week — bookmark them if you're building on Claude in production.
Rate Limit Error — Too Many Requests
You've exceeded RPM, TPM, or ITPM limits for your usage tier. Fix with backoff, tier upgrade, or batching.
Overloaded Error — Retry With Backoff
Anthropic's servers are temporarily overloaded. Not your fault — but your retry logic needs to handle it gracefully.
Context Window Exceeded — Prompt Too Long
Your prompt plus expected output exceeds the model's context limit. Fix via chunking, summarization, or model swap.
Authentication Error — Invalid API Key
API key missing, malformed, or revoked. Check environment variables and header format.
Tool Use Failed Validation
Your tool schema doesn't match Claude's requirements. Common with missing required fields or type mismatches.
Streaming Connection Dropped
SSE stream terminated mid-response. Fix via resumable streaming, client-side reconnection, or non-streaming fallback.
Prompt Caching Cache-Control Invalid
Your cache_control block placement or format is rejected. Common when adding caching to existing message arrays.
Extended Thinking Budget Error
budget_tokens value invalid or exceeds max_tokens. Common when enabling reasoning on existing calls.
Anthropic Claude Rate Limits
Rate limits are the #1 cause of Claude API errors in production. Anthropic uses a tier system based on cumulative spend. Limits apply per-model and reset every minute (RPM/TPM) or every day (RPD/ITPM).
| Tier | Qualification | Claude Opus (RPM) | Claude Sonnet (RPM) | Claude Haiku (RPM) |
|---|---|---|---|---|
| Tier 1 | Deposit $5, wait 7 days | 50 | 50 | 50 |
| Tier 2 | $40 cumulative spend | 1,000 | 1,000 | 1,000 |
| Tier 3 | $200 cumulative spend | 2,000 | 2,000 | 2,000 |
| Tier 4 | $400 cumulative spend | 4,000 | 4,000 | 4,000 |
| Enterprise | Contact sales | Custom | Custom | Custom |
Actual limits change frequently. Verify against your Anthropic Console rate limits page for current numbers. TPM and ITPM (Input Tokens Per Minute) limits also apply separately.
Claude Model Catalog
Claude offers three model families, each optimized for different tradeoffs between intelligence, speed, and cost. Choose based on your latency budget and reasoning requirements.
| Model | Context Window | Best For | Input $/M tokens | Output $/M tokens |
|---|---|---|---|---|
| claude-opus-4 | 200K tokens | Highest reasoning, complex coding, research | $15 | $75 |
| claude-sonnet-4.5 | 200K tokens | Balanced performance, coding, agentic workflows | $3 | $15 |
| claude-haiku-4.5 | 200K tokens | Fast, cheap, high-volume tasks, latency-sensitive | $1 | $5 |
Model versions, context windows, and pricing update frequently. Verify against official Anthropic pricing.
Browse Errors by Category
Claude errors organized by what functionally went wrong. Handy when you know the symptom but not the specific error code.
All Documented Claude Errors
Filterable list of every error we've documented. Search by error code, symptom, or keyword.
Showing 15 of 254 documented errors. More errors are being added weekly.
Frequently Asked Questions
By a wide margin, rate_limit_error (HTTP 429) is the most-searched Claude error. It affects nearly every developer at some point, especially during Tier 1 usage or when scaling agentic workflows. The fix depends on cause — see our rate limit error page for the full breakdown.
Yes. Rate limits are per-model within your tier. Hitting a rate limit on Opus doesn't affect your Sonnet or Haiku limits. This lets you fall back to smaller models when Opus limits are exhausted. See the rate limits table above for typical numbers.
Mostly the same error taxonomy, but wrapped in AWS or GCP error envelopes. Bedrock errors come through ThrottlingException and ValidationException, while Vertex uses gRPC status codes. The underlying causes are identical. We flag Bedrock/Vertex-specific quirks on each error page when they apply.
Check three sources in order: (1) our Anthropic status page which pings the API every 60 seconds, (2) the official Anthropic status page, and (3) whether errors are 5xx (Anthropic's problem) vs 4xx (your problem). If you see overloaded_error or api_error across multiple retries, it's likely upstream.
Yes. Every error page includes fixes for Python, TypeScript, and curl. The error semantics are identical across SDKs — only the syntax differs. We test all three code paths on every error page before publishing.
We prioritize by search volume, community demand (Discord/Reddit signal), and internal encounters. The 15 errors visible here are the highest-searched Claude errors on Google. Deeper coverage of the remaining 239 errors rolls out over the next 3 months. See our Methodology for the full framework.
Yes — it's been generally available since 2025 and is used in production by major companies. But it changes more frequently than the standard API, and pricing implications (thinking tokens count separately) trip up new users. Read our Extended Thinking Budget Error page before enabling it.
Email info@aierrorhub.com with subject "New Error Report" and include the exact error message, the endpoint you were calling, and the SDK version. We add reader-reported errors within 5 business days if we can reproduce them.