Repository navigation
Cache write tokens not accounted for when using OpenRouter, causing cost underestimation #18440
Description
Description
When a cache miss occurs via OpenRouter, opencode displays a near-zero cost even though the actual billing is much higher. This causes the in-app cost tracker to be way off from what OpenRouter bills.
The root cause: cacheWriteInputTokens falls back to 0 for OpenRouter because the code only looks for cache write tokens in anthropic and bedrock metadata. But OpenRouter actually exposes this in the usage response:
"usage_response": {
"prompt_tokens_details": {
"cached_tokens": 10318,
"cache_write_tokens": 0
}
}The fix would be to also read input.usage?.cacheWriteInputTokens (or whatever field name the vercel/ai SDK maps cache_write_tokens to) as a fallback:
const cacheWriteInputTokens = safe(
input.metadata?.["anthropic"]?.["cacheCreationInputTokens"] ??
input.metadata?.["bedrock"]?.["usage"]?.["cacheWriteInputTokens"] ??
input.metadata?.["venice"]?.["usage"]?.["cacheCreationInputTokens"] ??
input.usage?.cacheWriteInputTokens ?? // OpenRouter
0
)The exact field name on input.usage needs to be verified against how the vercel/ai SDK maps prompt_tokens_details.cache_write_tokens.
Plugins
N/A
OpenCode version
1.2.27
Steps to reproduce
- Use Claude sonnet 4.6 with openrouter
- Create a 50k token conversation
- wait 10 min (TTL is 5 min)
- send a new message
- See that price estimation only increases marginally (<0.1$) but the real cost is >0.1$
Screenshot and/or share link
Don't have it there.
But 4$estimated, 20$ real
Operating System
No response
Terminal
No response
Activity
iandreyshev commentedon Mar 24, 2026
We're hitting the same issue with LiteLLM proxy routing to AWS Bedrock (bedrock/us.anthropic.claude-sonnet-4-6).
LiteLLM correctly returns cache_creation_input_tokens and cache_read_input_tokens in the usage object:
But since LiteLLM uses the OpenAI-compatible format, there's no metadata["bedrock"] or metadata["anthropic"] in the response. OpenCode only reads cache write tokens from provider-specific metadata, so cacheWriteInputTokens is always 0.
In our case, OpenCode shows ~$0.03 while LiteLLM reports ~$0.07 for the same session. The difference comes from 10798 cache_creation tokens priced at $3.00/M (regular input) instead of $3.75/M (cache write).
The fallback to input.usage.cacheWriteInputTokens proposed above would fix this for all OpenAI-compatible proxies.
Count cache writes, not just cache hits