--- title: "Track AI agent cost per session, model and repo" description: "Tokens and cost per session, agent, model, repo and day, your most expensive sessions, and where each number comes from." canonical: https://coding.sh/docs/cost human_url: https://coding.sh/docs/cost markdown_url: https://coding.sh/docs/cost.md agents_url: https://coding.sh/agents.md llms_url: https://coding.sh/llms.txt updated: 2026-11-11 --- # Track cost and tokens Tokens and cost per session, agent, model, repo and day, your most expensive sessions, and where each number comes from. coding.sh reads the token counts that any AI agent records, turns them into cost, and keeps both with each session. This page gets your totals, finds the most expensive sessions, and explains where each dollar figure comes from. ## Before you start - coding.sh is installed and has sessions ([Quickstart](https://coding.sh/docs/quickstart)). - To count sessions from before you installed, run `coding.sh ingest --backfill --since 30d` once. ## 1. See totals ```sh coding.sh usage --since 7d ``` ```text AGENT SESSIONS TURNS TOKENS OUTPUT TOK/TURN CACHE READ COST claude-code 29 108 9.8M 86.9k 6899 93% ~$6.92 codex 14 38 4.0M 52.4k 26.1k 76% ~$2.08 opencode 5 5 1.1M 8542 26.5k 89% ~$0.7937 coding-sh-triage 8 8 13.4k 221 1679 0% ~$0.0095 ~ = includes cost estimated from tokens × list price (no reported cost). Cache breaks: 32, 349.9k prompt tokens billed again (~$0.5240 extra). Causes per session: coding.sh show . ``` `coding-sh-triage` is an SDK app; its rows sit next to the coding agents. `--since` defaults to `30d`. `-a codex` keeps one agent, and `--where` takes any [query](https://coding.sh/docs/query). ## 2. Find the most expensive sessions ```sh coding.sh sessions --where 'sort:-cost limit:5' ``` ```text ID AGENT STARTED DURATION TURNS TOOLS TITLE e2ac1139-922a-46bd-8350-bcce0f43da68 claude-code Nov 5 02:36 3m41s 4 11 Stream large blobs instead of buffering them 842f85fe-a932-4e47-8ed2-de9b968eaa72 claude-code Nov 8 18:53 4m01s 3 11 Fix the flaky hook test on Windows c9adf6ea-3327-47ed-8b99-62f696ca7bc7 claude-code Nov 7 21:33 4m25s 3 11 Handle an empty spool in ingest ea916f85-7248-4f70-8159-52d28269c6c5 claude-code Nov 6 21:22 3m24s 5 11 Fix the N+1 query for subagents 4b3edd8a-56c9-4cb5-83d4-213c0706f973 claude-code Nov 10 02:08 2m35s 4 6 Add a --json flag to coding.sh show ``` The list has no cost column. To see the figures, ask for JSON: ```sh coding.sh sessions --where 'sort:-cost limit:5' --json --fields id,title,stats.usage.costUsd ``` `usage --json` also carries the five biggest sessions of the window in `topSessions`, ranked by cost when any session has one, else by output tokens (`topSessionsRankedBy` says which). Narrow the list with the same query: `agent:codex`, `repo:coding-sh`, `since:24h` or `cost>$0.25`. Subagents are left out unless you pass `--subagents`. ![The session list in the viewer, newest first, with agent, tokens and cost per row](https://coding.sh/_astro/sessions-list.CIcIRzY2_awDMF.webp) coding.sh ui: the session list ## 3. Open one ```sh coding.sh show e2ac1139 ``` ```text Tokens in 47 out 5.2k cache 547k $0.3823 (estimated) Cache breaks 1: 2.9k prompt tokens billed again, $0.0100 extra (other 1) ``` The `Tokens` line holds input, output, cache reads and cost. Each model call below it shows its own cost, with `est` when it is an estimate. Long sessions usually spend their tokens re-reading context: look for a high cache-read share, a `compaction` flag and repeated tool calls. `coding.sh ui` opens the same session as a timeline. ## 4. Split by agent, model, repo or day ```sh coding.sh usage --since 7d --by model ``` ```text MODEL SESSIONS TURNS TOKENS OUTPUT TOK/TURN CACHE READ COST claude-sonnet-4-6 28 99 9.7M 84.9k 8009 93% ~$6.51 gpt-5 14 38 4.0M 52.4k 26.1k 76% ~$2.08 claude-opus-4-8 4 13 1.1M 9773 5796 94% ~$1.17 claude-haiku-4-5 5 4 74.0k 932 4180 60% ~$0.0404 gpt-5-mini 5 5 6079 109 1216 0% ~$0.0017 ``` | `--by` | One row per | | --- | --- | | `agent` | Agent or app (the default) | | `model` | Main model | | `agent-model` | Agent and model pair | | `repo` | Repo root | | `day` | Local day, newest first | ## Where cost comes from Transcripts and most OpenTelemetry spans carry tokens, not dollars. Each model call's cost comes from the first of these that exists: - Billed. The bill from an LLM gateway, after `coding.sh gateway enrich` ([LLM gateways](https://coding.sh/docs/ai-gateway)). - Provider-reported. A cost the agent or app records: Claude Code runs that coding.sh starts in evals and experiments (`total_cost_usd`), OpenCode, Pi, or a span with `gen_ai.usage.cost` or `cost_usd`. - Estimate. Each kind of token times the model's list price, marked `~`. A model with no price keeps no cost, never $0. `usage` names it under the table, with the fix: a `[prices.""]` table in config.toml ([Configuration](https://coding.sh/docs/config#prices)). On a flat subscription, an estimate is what the same tokens cost at API list prices, not what you paid. Which agents record tokens | Agent | Tokens | | --- | --- | | SDK and OpenTelemetry apps | Every model call the app traces | | Claude Code, Codex, OpenCode, Pi, Factory Droid, Amp, T3 Code | Every session | | Antigravity CLI, Goose, GitHub Copilot CLI | From their history files, imported by `coding.sh ingest --backfill`; their hooks carry none | | Cursor CLI | None, so no cost either | | `coding.sh run -- ` | None, except Aider's token and cost lines in eval runs | When prices change, estimates refresh on their own. To recompute them now: ```sh coding.sh usage --since 7d --reprice ``` It prints `Repriced N model calls in M sessions` on stderr, then the table. Rates, tiers and Claude Code's own tally The price table covers current Claude, OpenAI, Gemini and common open models. A model with no row is priced like its nearest sibling in the same family and major version, shown in `payload.pricedAs`. Each call is priced at the rate it 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 does not reprice the session. Claude 4.6 and later have no long-context surcharge. Codex rollouts record no tier, so their calls get standard rates. A call priced at another rate says so in `payload.priceTier` (`batch`, `flex`, `long-context`, or a mix such as `batch+long-context`). Totals by tier are in `estimatedCostByTier`. Claude Code also keeps its own tally at the end of its transcript. A backfill stores it on `session.end` as `payload.reportedUsage`. It is never added to the session's cost; compare the two to check your prices. ## Prompt cache health `coding.sh usage --since 7d --cache` shows cache hit rate, reads, writes, breaks and the cost of each break per agent; `coding.sh cache doctor` grades it and lists the causes. ```text AGENT SESSIONS CACHED READ WRITE R:W BREAKS WASTE claude-code 29 93% 9.0M 657.6k 13.71:1 13 ~$0.1808 codex 14 76% 3.0M 0 - 18 ~$0.3305 opencode 5 89% 979.4k 0 - 1 ~$0.0128 ``` A break is a model call whose cached prompt prefix was gone, so the prompt was billed again. To find the worst sessions: `coding.sh sessions --where 'cache.breaks>0 sort:-cache.cost'`. ## Read the result | Mark | Means | | --- | --- | | `~$6.92` | Part of the figure is an estimate at list price. | | `≥$1.20` | A lower bound: some sessions in the row have tokens but no price. | | `-` | No session in the row has a price. | | `est` | This model call's cost is an estimate (`show`). | | `(estimated)` | The session's cost includes an estimate (`show`). | | `estimatedCostUsd` | The estimated part, in `--json` and the API. | | `TURNS` | User turns of top-level sessions. | | `TOK/TURN` | Input, output and cache-write tokens per turn, top-level sessions only. | | `CACHE READ` | Cache reads as a share of all prompt tokens. | ## If something goes wrong | Symptom | Check | | --- | --- | | No `COST` column | No session in the window has a price. Look for model names under the table and add `[prices.""]`. | | An agent shows tokens but `-` for cost | Its model has no price. Add one, then run `coding.sh usage --reprice`. | | An agent shows no tokens | Some agents record none in their hooks; see "Which agents record tokens" above. Run `coding.sh ingest --backfill` to read their history files. | | Totals look low | Sessions from before you installed are missing. Run `coding.sh ingest --backfill --since 30d`. | | The cost differs from your invoice | It is an estimate at list price. Route calls through a gateway and run `coding.sh gateway enrich` for billed cost. | ## Next steps - [Find and read runs](https://coding.sh/docs/explore): read one session step by step. - [Capture every model call with the proxy](https://coding.sh/docs/how-to/capture-with-proxy): cost and cache use per model call, and why a cache missed. - [coding.sh usage](https://coding.sh/docs/cli/usage): every flag. - The MCP `usage` tool returns the same totals to your agent ([MCP server](https://coding.sh/docs/mcp)). To set coding.sh up for a user, follow the agent workflow at [/agents.md](https://coding.sh/agents.md).