Estimate

Before launching a full anatoly run, you almost always want to know two things: how much will this cost and how long will it take. The estimate command answers both without making a single LLM call.

anatoly estimate

Token counts come from a local tiktoken pass over the scanned files, prices are read from an on-disk cache (litellm + OpenRouter), and the wall-clock estimate uses calibrated medians from your previous runs. If no prior scan exists, estimate runs one automatically.

When to use it#

• Before your first run on a new repository — to size the operation and pick the right billing mode.
• Before wiring it into CI — anatoly estimate --json emits a stable, versioned payload you can budget against.
• After tweaking config — concurrency, RAG mode, triage thresholds, or model overrides all change the forecast.
• As a sanity check — if total billed shows an unexpected number, your provider mode or model resolution is misconfigured.

Reading the output#

The rendered view is built bottom-up so the verdict (Forecast) lands closest to the prompt:

  Project Info
  ──────────────
  name         @r-via/anatoly
  version      0.9.5
  languages    TypeScript 97% · JSON 2% · Shell 2%
  frameworks   Commander · Vitest
 
  Configuration
  ──────────────
  concurrency   8 files · 24 Claude slots
  cache         on
  rag           advanced — 166 files · 888 fns · 2897 chunks
  docs          0 changed, 34 cached · project 21 changed
 
  Cost breakdown based on latest public provider price
  ──────────────
  category       step               cost   mode           model
  axis           best_practices   $53.46   subscription   anthropic/claude-sonnet-4-6
                 correction       $27.34   subscription   anthropic/claude-sonnet-4-6
                 ...
  deliberation                    ~$3.54   subscription   anthropic/claude-opus-4-6
  summary                          $1.06   subscription   anthropic/claude-haiku-4-5
  embed          code              $0.00   local          nomic-embed-code Q5_K_M (local)
                 text              $0.00   local          Qwen3-8B Q5_K_M (local)
 
  total billed                     $0.00
  consumption                   ~$122.41
 
  Forecast
  ──────────────
  files    185 of 190  (5 skipped by triage)
  tokens   ~11.1M in / ~7.1M out + ~418K embed
  cost     $0 in subscription mode  (ensure quota for ~$122.41)
  time     ~1h 19m  (calibrated)

Project Info — name, version, detected languages, frameworks.

Configuration — runtime settings and indexing scope. The rag line carries the mode and indexing breadth; the docs line shows how many internal docs are changed vs cached and how many project files changed since the last run.

Cost breakdown — one row per pipeline step that hits the LLM (or embedding API), grouped by category (axis → deliberation → summary → embed → internal-doc). The mode column is the key one to read:

• subscription — covered by your Claude Code OAuth quota; you don't pay this directly.
• api — real per-token billing.
• local — free local runtime (e.g. local embedding models).

Two totals close the table:

• total billed — what actually leaves your wallet (sum of api-mode rows).
• consumption — the API-equivalent magnitude across all rows. Useful to size your subscription quota.

Forecast — the headline. The cost line is mode-aware:

• $0 in subscription mode (ensure quota for ~$X) when fully covered by OAuth.
• $X in consumption mode when fully API-billed.
• $X billed (~$Y consumption equivalent) for mixed setups.

The time line is tagged (calibrated) once you have prior runs to learn from, otherwise (default).

JSON mode#

anatoly estimate --json

Emits a versioned payload (schemaVersion: 1) to stdout, with logs redirected to stderr and the banner suppressed. The shape mirrors the rendered table one-for-one.

For the full schema and an annotated example, see CLI Reference → estimate.

Caveats#

• The cl100k_base tokenizer is Claude-compatible but not identical — actual costs typically fall within 10–20% of the estimate, depending on prompt caching and retry rates.
• Time estimates are best after a few real runs have populated the calibration cache; the first estimate on a fresh repo uses defaults.
• estimate reads the same config as run, so make sure --config (or the implicit .anatoly.yml) points to the file you actually intend to use.