CodeGraph made local token counts spike, but provider logs showed the run used fewer tokens
TruthTellerTom · reddit · 2026-07-28
CodeGraph can make local token meters look worse while provider logs show savings
A Reddit user compared the same planning prompt in OpenCode/OpenChamber with and without CodeGraph. Locally, the run with CodeGraph appeared to use far more tokens, but the provider's own usage logs showed the opposite: the CodeGraph run actually saved tokens.
The post raises a practical question for agent tooling:
- local context meters may only reflect what the client sees
- provider logs can tell a different story
- faster runs don't necessarily imply higher billed usage
- repeated prompts may also interact with caching or internal routing
The thread is essentially about how to interpret token accounting when using coding-agent stacks.
More from coding & agent
- Gemini CLI fixes macOS sandbox crashes by embedding Seatbelt profiles — amelidev · 2026-07-28
- Nativ v0.1.0 Released: Run AI Models Locally on Apple Silicon with Multi-IDE Support — max_paperclips · 2026-07-28
- A practical guide to when to use ChatGPT, Claude, and agent modes — Simon Willison · 2026-07-28
- Locally Hosted Kimi K3 Beats Cloud GPT-5.6 in 3D Physics Generation — rohanpaul_ai · 2026-07-28
- Hermes Desktop Avatar wraps a local agent in an always-on-top chat UI — TinfoilTricorn · 2026-07-28
- Berkeley-led paper maps three autonomy levels for software development — dawnsongtweets · 2026-07-28