Environment

• Platform (select one):
  • Anthropic API
  • [XX] AWS Bedrock
  • Google Vertex AI
  • Other:
• Claude CLI version: 1.0.3
• Operating System: macOS 24.4.0 (Darwin) on ARM64 architectur
• Terminal: VS Code

Bug Description

Prompt caching works properly on Claude 3.7 Sonnet but does not work on newer Claude models (Claude Sonnet 4 and Claude Opus 4).

Steps to Reproduce

1. Start a Claude Code session with default model (Claude 3.7 Sonnet)
2. Note that cache writes are properly recorded
3. Switch to Claude Sonnet 4 or Claude Opus 4 using model command
4. Observe that cache read/write counts remain at 0

Expected Behavior

Claude 4 models should properly utilize prompt caching to reduce token usage and costs.

Actual Behavior

Only Claude 3.7 Sonnet shows cache activity (both reads and writes). Claude Sonnet 4 and Claude Opus 4 show 0 cache read/write regardless of

Additional Context

Edit: Separately, the ANTHROPIC_SMALL_FAST_MODEL does not use prompt caching regardless of what model is set. This means that Claude 3.7 does not use prompt caching when ANTHROPIC_SMALL_FAST_MODEL is set to Claude 3.7. This is clearly a bug considering that Claude 3.7 prompt caching does work when it is set as the main model (ANTHROPIC_MODEL).

Verified with a clean install and completely clean configs.

Anthropic guys, I am begging you to fix this. It makes the Claude 4 family completely unusable for coding - the costs are astronomical compared to using Claude 3.7 with prompt caching - and worse - the latency is terrible once you have a bit of context, especially with Claude 4 Opus!

Thanks in advance - love the tool - life changing!