Skip to content

fix(pricing): Kimi K3 Bedrock rates + cost fallback for omp/pi via LiteLLM - #192

Merged
aarora79 merged 1 commit into
aarora79:mainfrom
shekharprateek:fix/kimi-k3-bedrock-pricing
Sep 24, 2026
Merged

aarora79 merged 1 commit into
aarora79:mainfrom
shekharprateek:fix/kimi-k3-bedrock-pricing

Conversation

@shekharprateek

Copy link
Copy Markdown

Summary

  • Add moonshotai.kimi-k3 to bedrock_pricing.py with Global CRIS Standard tier rates ($3.00/$15.00 per 1M input/output, $0.30/$3.75 cache read/write), sourced from the Bedrock pricing page on 2026-09-23.
  • When an agent reports no cost (omp and pi report $0 through the LiteLLM proxy because it doesn't relay Bedrock pricing), _metrics_from_result now falls back to bedrock_pricing.cost_usd -- the same lookup the codex path already uses. Models not in the table still get null, so vLLM runs (hardware-derived cost) are unaffected.

Test plan

  • uv run python -m unittest tests.test_bedrock_pricing passes (62 tests)
  • Verify _normalize("moonshotai.kimi-k3") resolves to the price table key
  • Existing agents that report their own cost (claude, kiro) are unaffected -- the fallback only fires when total_cost_usd is None

…n agent reports no cost

Two fixes so omp-via-LiteLLM runs record a real dollar cost instead of null:

1. Add moonshotai.kimi-k3 to bedrock_pricing.py ($3.00/$15.00 per 1M
   input/output, Global CRIS Standard tier, sourced from the Bedrock
   pricing page on 2026-09-23).

2. In _metrics_from_result, when the agent reports no cost (omp and pi
   report $0 through the LiteLLM proxy because it doesn't relay Bedrock
   pricing), fall back to bedrock_pricing.cost_usd -- the same lookup
   the codex path already uses. Models not in the table still get null,
   so vLLM runs (hardware-derived cost) are unaffected.
@aarora79
aarora79 merged commit 6a4a430 into aarora79:main Sep 24, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants