How to Check opencode Token Usage and Cost

opencode is an open-source terminal AI agent built for developers who want full control over their model configuration. It supports multiple providers — Anthropic, OpenAI, AWS Bedrock, and more — and keeps a detailed JSON log of every session it runs. That log is exactly what the ccusage library (which powers whoburnedmore) knows how to read. Unlike a closed IDE like Cursor, opencode's transparency means a single command can give you token-level accuracy: which models you hit, how many tokens each session consumed, and what each session cost at current API rates.

Where does opencode store its usage logs?

opencode writes session logs to a local directory under your home folder. The exact path varies slightly by platform and version, but the canonical location is:

# macOS / Linux (XDG-compliant default)~/.local/share/opencode/ # Each session creates a file like:~/.local/share/opencode/sessions/2026-06-14T09-30-00_session.json # Quick check: how many sessions do you have?$ ls ~/.local/share/opencode/sessions/ | wc -l     47

Each session file contains the full message history, model identifiers, and token usage broken down by input and output for each turn. The ccusage library reads these files directly — no API calls required — so the count is exact, not estimated from request metadata.

What if the directory doesn't exist?

If ~/.local/share/opencode/sessions/is empty or missing, either opencode hasn't been run yet on this machine, or it was installed to a non-default location. You can check opencode's configuration file for a custom dataDir setting. If you used a package manager that installs to a different prefix, find ~ -name "*.opencode*" -type d 2>/dev/null will locate it.

How do I run the token total command?

No installation or account required. The command uses npx to pull the latest version every time:

$ npx whoburnedmore↳ scanning local usage logs…  found opencode sessions: 47   TOOL          TOKENS       MODEL SPLIT         EST. COST  opencode      11.4M        claude-3-5: 72%     $33.10                             gpt-4o:     28%  claude code    8.9M        claude-3-5: 100%    $29.60  codex          4.2M        gpt-4o:     100%    $11.40  ─────────────────────────────────────────────────────  total         24.5M                            $74.10

Notice the model split column: because opencode lets you switch providers mid-workflow (e.g. starting with Claude 3.5 Sonnet and falling back to GPT-4o for certain tasks), whoburnedmore tracks token counts per model within opencode so the cost estimate uses the correct per-token price for each.

How is opencode token cost calculated?

opencode supports multiple providers and model tiers, so the cost formula aggregates across all of them. For a given session, whoburnedmore calculates:

Where m ranges over each model used in the session, in and out are the input and output token counts for that model, and p is the published per-token price. The session costs are then summed across all sessions for the total. Input tokens are always cheaper than output tokens — on Claude 3.5 Sonnet, for example, input runs at $3/M versus $15/M for output, so a prompt-heavy workflow costs noticeably less than a generation-heavy one.

Can I see opencode usage by day or project?

Yes — pass the --breakdown flag to get daily totals:

$ npx whoburnedmore --breakdown=daily↳ filtering to opencode sessions…   DATE         TOKENS    SESSIONS    COST  2026-06-14   2.1M      8           $6.20  2026-06-13   1.8M      6           $5.30  2026-06-12   3.4M      11          $9.90  2026-06-11   0.6M      2           $1.80  7-day avg    1.6M/day   5.4/day     $4.80/day

Comparing opencode to other terminal agents

One of the most useful things the daily view reveals is how your opencode usage compares to other terminal agents on the same day. If you switch between opencode and Claude Code depending on the task, the breakdown shows you which tool you actually rely on more:

Is reading my opencode logs safe?

whoburnedmore reads the metadatain opencode's session files — token counts, model names, timestamps — and ignores the message content entirely. Your prompts, the code you pasted in, the files opencode read during a session — none of that is parsed or transmitted. Only aggregate numbers (total tokens, estimated cost per session) are ever sent to the leaderboard, and only if you choose to submit them. Run npx whoburnedmore --local to keep everything on your machine and open the dashboard in your browser instead of submitting to the public board. 🛡️

If you also use newer AI coding agents like Amp, Droid, or Goose alongside opencode, the guide to checking Amp, Droid, and Goose usage covers the same one-command workflow for those three tools, with tool-specific log locations and gotchas for each.

1. 1

   Run the command (30 seconds)

   npx whoburnedmore — it finds your opencode sessions automatically. No config file to edit.
2. 2

   Check the model split

   If you mix Anthropic and OpenAI models in opencode, confirm the cost estimate reflects both rate schedules. The model column in the output shows you which models were hit.
3. 3

   Bookmark the private dashboard

   Run npx whoburnedmore --local to open a browser dashboard that stays entirely on your machine. Useful for day-to-day monitoring without publishing to the leaderboard.