Spend less tokens | Friends of the Crustacean 🦞🤝 | Page 1

sleek gull Apr 28, 2026, 3:59 PM

#

My assistant spent 32 million tokens in 2 days.
Is that normal and if not can someone help me find out why it spent that much and how to reduce it?
My assistant answers mainly question and does tasks e.g. setup in notion, daily reminder.
I host on hostinger and use an openai api

crystal wadiBOT Apr 28, 2026, 3:59 PM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

calm pondBOT Apr 28, 2026, 4:03 PM

#

32M tokens in 2 days is high for mostly Q&A + reminders. Not impossible if it’s carrying huge context/tool outputs, but worth investigating.

First, please paste:

openclaw --version
openclaw status --all
openclaw status --usage

And in the assistant chat run:

/usage full
/usage cost
/context detail

#

Don’t paste API keys/tokens.

Likely causes:

one long session repeatedly re-sending a huge context
Notion/browser/tool calls returning large pages/results
cron/heartbeat jobs firing too often
subagents or retries looping
media/screenshots/transcripts
OpenAI key being used by something else too

Quick reductions:

run /compact in long chats
start fresh sessions for unrelated tasks
reduce/disable frequent cron/heartbeat jobs
trim big tool outputs/workflows
use a cheaper/smaller model for reminders/background work; for current model picks, ask in #models
put OpenClaw on its own OpenAI project/API key and set usage limits in OpenAI

#

OpenClaw has built-in usage views via /status, /usage full, /usage cost, and openclaw status --usage. (docs.openclaw.ai) OpenAI’s usage dashboard can filter by project/model/API capability, and OpenAI also supports org/project usage limits. (help.openai.com)

sleek gull Apr 28, 2026, 4:30 PM

#

calm pond Don’t paste API keys/tokens. Likely causes: - one long session repeatedly re-se...

I have one chat I use for everything in telegram. But its all related.
Here are the usage results:
Session Usage
Input: 158.9k tokens
Output: 910 tokens
Total: 158.9k tokens
Context: 79% of 200k
Model: gpt-5.4

Agents: 1 · 1 bootstrap file present · sessions 2 · default main active 11m ago
Usage: no provider usage available

calm pondBOT Apr 28, 2026, 4:30 PM

#

Yep, that explains it: 158.9k input tokens per reply is very large.

At that rate, 32M tokens ≈ ~200 replies/tool loops in 2 days. So the spend is probably not a mystery leak — it’s your one Telegram session carrying a huge context every turn.

What I’d do now:

Run:
```
/context detail
```
This should show what’s filling the 79% context — chat history, tool results, bootstrap file, memory, etc.
Run:
```
/compact
```
That’s the fastest fix for the current chat.

#Spend less tokens