Description

Since v1.3.4 (which includes AI SDK v6 via c33d9996f feat: AI SDK v6 support (#18433)), token usage total is double-counted for Anthropic/Bedrock providers. The total field reports roughly 2x the actual token count, and tokens.input (the non-cached portion) is inflated.

Root cause: AI SDK v6 normalized inputTokens to include cached tokens for all providers (PR vercel/ai#10975). Previously, Anthropic excluded cache from inputTokens. The excludesCachedTokens flag in getUsage() still assumes the old v5 behavior — it skips subtracting cache tokens for Anthropic, then adds them back when computing total, effectively counting cache twice.

Evidence from real sessions (v1.3.4, Anthropic direct):

input: 139499, cache.read: 139417, cache.write: 81, total: 279057

input - cache.read - cache.write = 1 (only 1 non-cached token), proving inputTokens now includes cache. But total = 139499 + 60 + 139417 + 81 = 279057 — cache counted twice. Correct total should be 139559.

This pattern is consistent across all requests: input ≈ cache.read + cache.write + 1.

Plugins

None

OpenCode version

v1.3.4 / v1.3.5

Steps to reproduce

1. Use Anthropic provider directly with prompt caching enabled
2. Observe tokens in session usage data
3. total ≈ 2x expected value; input includes cached tokens but code doesn't subtract them

Screenshot and/or share link

N/A — data-level issue, not visual

Operating System

macOS

Terminal

OpenCode Desktop