Token count is way off

[abdurahmanshiine]

Description

Hey there,

I started using OpenCode today and I got an out-of-credit error very quickly, that surprised me because the token count in OpenCode said: ~85k tokens, but when I went to the billing page of my API provider (Fireworks AI) it said I had used close to 3.55m tokens, which is a crazy gap. And so I'm wondering where is that difference coming from

OpenCode version

1.0.185

Steps to reproduce

No response

Screenshot and/or share link

Operating System

WSL

Terminal

Windows Terminal

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Incorrect token usage #5440: Incorrect token usage - reported discrepancy between token count display and actual API usage
• bug: token cost seems inaccurate #2020: bug: token cost seems inaccurate - reported incorrect cost calculations and large discrepancies in token counts
• web and tui reports token usage differently. #3322: web and tui reports token usage differently - reported discrepancies in how token usage is calculated/reported between different interfaces
• Ridiculously big usage of input tokens for simple tasks #254: Ridiculously big usage of input tokens for simple tasks - reported unexpectedly high token usage for simple operations

Feel free to ignore if none of these address your specific case.

[abdurahmanshiine]

Taking a look at the stats I got this, which is really weird because I have only sent 12 messages (and not 75)

[rekram1-node]

and u only used fireworks in opencode? that usage couldn’t be from elsewhere?

That big of a gap just doesnt make sense UNLESS the session that this happened in had a bunch of subagent calls?

Rn it only shows the parent totals instead of including subagent costs and token usage

[abdurahmanshiine]

and u only used fireworks in opencode? that usage couldn’t be from elsewhere?

Yes, I actually created one Fireworks account specifically to use it with OpenCode, and I didn't use any subagents (I didn't even know how to invoke them until I checked now 😅 )

Not sure if it changes anything, but most of the messages in the convo (~10 of them) were using the planning mode, and only 2-3 used building mode

[coreyward]

Rn it only shows the parent totals instead of including subagent costs and token usage

@rekram1-node Yikes. That caught me off guard as well. I set up a brand new key to use OpenCode with Claude Opus 4.5 via API. I sent a single prompt/request in plan mode and it used a bunch of subagents to gather context. The TUI reported this as $0.16 of usage. The Anthropic billing dashboard, however, reported as $2.14. That's 13× the reported price.

[yassine-safraoui]

I ran into a similar bug, but it's a little more specific. I was in a thread and decided to fork it from a certain message, then I started using a different model in the new fork. As I kept using the fork thread, I noticed the total price actually started decreasing. It looks like the math is flipped—instead of adding the cost of the new messages to the total, the system is performing a minus operation, so the more I talk, the "cheaper" the thread gets. Definitely seems like there's a bug in how the cost is being calculated, not sure what the root cause of. this is.