opencode-model-cost

Live per-model tokens, cost, and TPS in the OpenCode TUI sidebar.

Shows the model split of the current session tree in the right sidebar, one model per line, updated in real time:

Models  $2.19  135k tok
glm-5.3  89k  $2.10  ·~12.3tps
glm-5.3-flash  12k  $0.09
local  34k

The heading shows the totals; each line shows the model, its cumulative token count (input + output + reasoning), its cumulative cost (hidden when $0), and its tokens-per-second rate — the last run's rate is retained between runs, and the currently streaming model's line is highlighted.

Why

OpenCode's built-in stats show totals per session. Once you run multiple models — a flagship main agent delegating to cheap flash subagents, or a free local lane — you lose sight of who is actually spending what. This plugin breaks the session tree down per model, live, so you can see the split while you work.

Requirements

• OpenCode >= 2.0.24 (use plugin 0.2.x for OpenCode v1)
• The TUI (the plugin does nothing in CLI/web mode)

Install

Add the plugin to your global CLI config (~/.config/opencode/cli.json):

{
  "plugins": ["opencode-model-cost"]
}

Or install via the CLI:

opencode plugin add opencode-model-cost

Then restart OpenCode.

How it works

• Reads every assistant message in the viewed session's tree — the session itself plus all subagent child sessions (attributed via parentID)
• Groups cumulative tokens and cost per model
• Tracks streaming text deltas to compute a per-model live TPS (5-second rolling window, byte-based token estimate); the last run's rate stays visible and the streaming model's line is highlighted
• Renders a block in the right sidebar of the session view, in theme colors, one line per model

Limitations

• TPS is available only for responses streamed during the current TUI run; restored session history still includes all models' tokens and cost
• Token counts exclude cache reads/writes to keep the display readable; use opencode stats for full accounting
• TPS is an estimate based on streamed text/reasoning bytes (~5 bytes per token), marked with ~ in the display

License

MIT