Find your most expensive sessions

Rank sessions by cost or tokens, break spend down by repo, agent and model, and see where cost data comes from.

Rank sessions by cost or tokens, see where the spend goes by repo, agent and model, and open the worst one.

1. Totals by repo, agent and model

shell

coding.sh usage --since 30d                  # by agent (the default)
coding.sh usage --since 30d --by repo
coding.sh usage --since 30d --by agent-model
coding.sh usage --since 7d --by day

Column	What it counts
`TURNS`	User turns of top-level sessions (a subagent's prompt comes from its parent's turn)
`TOKENS`, `COST`	All tokens and cost, subagents included
`TOK/TURN`	Top-level input, output and cache-write tokens per user turn; no subagents, no cache reads
`CACHE READ`	Cache reads as a share of all prompt tokens

COST appears when at least one session in the window has a cost. ~ marks a cost that includes an estimate. ≥ marks a lower bound: some sessions in the row have tokens but no cost. A row with no priced session shows -.

Below the table, usage names the models with tokens but no price, and the fix: a [prices."<model>"] table in config.toml (Configuration).

2. The top sessions

usage --json includes the five biggest sessions in the window, ranked by costUsd when any session has a cost, else by outputTokens:

shell

coding.sh usage --since 30d --json | jq '.topSessionsRankedBy, .topSessions'

For a longer list, sort sessions --json yourself:

shell

coding.sh sessions --since 30d --limit 500 --json \
  | jq -r '.sessions | sort_by(-(.stats.usage.costUsd // 0), -(.stats.usage.outputTokens // 0))
           | .[:10][] | "\(.id)  \(.stats.usage.costUsd // "-")  \(.stats.usage.outputTokens // 0)  \(.title)"'

Add --agent codex to either command to narrow it, or --repo . to sessions for the current repo. sessions leaves subagents out unless you pass --subagents.

3. Open one

shell

coding.sh show <session-id>

The header has a Tokens line (input, output, cache reads, cost) and a Latency line with model call and tool timings. Long sessions usually spend their tokens on re-reading context: look for a high cache-read share, compaction flags, and repeated tool calls. coding.sh ui shows the same session as a timeline.

Where cost comes from

Agent transcripts carry tokens, not dollars, and so do most OpenTelemetry instrumentations. Each model call's cost comes from the first of:

the bill: calls through an LLM gateway (Vercel AI Gateway, OpenRouter, LiteLLM, Cloudflare, Portkey, Helicone) after coding.sh gateway enrich;
a cost the agent reports: Claude Code headless runs (total_cost_usd), OpenCode, Pi, or an OpenTelemetry span with gen_ai.usage.cost or cost_usd;
an estimate: input, output, cache-read and cache-write tokens times the model's list price.

Estimates are marked: ~ in usage, est on a call and (estimated) on a session in show, and estimatedCostUsd in --json and the API.

A model with no price row is priced like its nearest sibling in the same family and major version (gpt-6.2-sol as gpt-6.1-sol, shown in payload.pricedAs). A model with no price at all keeps no cost, never $0.

The built-in table covers current Claude, OpenAI, Gemini and common open models, with each provider's page and the date it was checked, in src/core/pricing.ts. Each call is priced at the rates it actually ran at:

Rate	Price
Long context	A prompt over OpenAI 272K, Gemini Pro 200K, Haiku 5.5 100K or Sonnet 4.5 and 4 (1M beta) 200K tokens
Batch	Half price at every provider
OpenAI Flex	Half price
OpenAI Fast (once called Priority)	About 2x
Gemini Priority	1.8x
Claude fast mode	Its own rates
Anthropic Priority Tier	Standard rates: it is a commitment, not a per-token price
Claude cache writes	1.25x input for the 5-minute cache, 2x for the 1-hour cache
US-only inference	1.1x on every token type for Claude 4.6 and later, on top of any other rate

Long context is decided per call, so one long call doesn't reprice the session. Claude 4.6 and later have no long-context surcharge. Where the rate comes from:

The service tier: Claude Code's usage.service_tier and usage.speed, its api_request log speed, an OpenTelemetry span's openai.response.service_tier (or gen_ai.*.service_tier), or AI SDK provider metadata. A response's tier beats the requested one.
1-hour cache writes: Claude Code's usage.cache_creation.ephemeral_1h_input_tokens, its api_request log's cache_creation_1h_tokens, or AI SDK provider metadata (payload.cacheWrite1hTokens).
US-only: inference_geo: "us" in Claude Code's usage (payload.inferenceGeo). Global routing, the default, is standard.
Codex rollouts record no tier, so their calls get standard rates; long context still applies.

A call priced at another rate says so in payload.priceTier (batch, flex, priority, fast, ultrafast, long-context, us-only, or a mix like batch+long-context), shown in show and the viewer's cost tooltip. Totals by tier are in estimatedCostByTier.

To match your bill, set your own prices, tiers included (Configuration). Estimates refresh when the table changes; coding.sh usage --reprice recomputes them now.

Claude Code also keeps its own tally, in a cost-state line at the end of the transcript. A backfill keeps it on session.end as payload.reportedUsage (costUsd, apiDurationMs, costUsdByModel, and unknownModelCost when it could not price a model). It is never added to the session's cost: compare the two to check your prices. The latest line wins.

Ask your agent

With the MCP server registered (coding.sh init does it), an agent can call the usage tool to get the same totals and top sessions, then session_get to read one.