@tensorcost/cli

Estimate and compare LLM prompt costs from the command line. Works offline with a bundled seed catalogue; when online, refreshes from TensorCost's public CloudFront snapshot (no API key required).

npx @tensorcost/cli estimate --prompt-file prompts/rerank.v3.txt \
                              --model anthropic/claude-3-sonnet
Estimated cost : $0.0042 / call
Input tokens   : 1,420
Output tokens  : 284
At 1,000 calls/day: $126.00/month

npx @tensorcost/cli compare \
      --baseline-file prompts/rerank.v2.txt \
      --candidate-file prompts/rerank.v3.txt \
      --model anthropic/claude-3-sonnet
baseline → candidate:
  Input tokens : 1,280 → 1,420  (+10.9%)
  Cost/call    : $0.0038 → $0.0042  (+10.9%)
  Verdict      : candidate is 10.9% more expensive per call.

Installation

npm install -g @tensorcost/cli
# or run without installing:
npx @tensorcost/cli <subcommand>

Both tensorcost and tc are registered as binary names, so after a global install you can use either:

tensorcost estimate --prompt "Hello" --model openai/gpt-4o
tc estimate --prompt "Hello" --model openai/gpt-4o

Commands

estimate

Estimate the cost of a single prompt invocation.

| Flag | Description | |------|-------------| | --prompt <text> | Inline prompt text | | --prompt-file <path> | Path to a prompt file | | --model <id> | Model identifier — see tensorcost models | | --provider <name> | Estimate the same prompt across all models for a provider (e.g. anthropic) | | --expected-output-tokens <n> | Override output token estimate (default: max(64, input/5)) | | --calls-per-day <n> | Also print a monthly cost projection | | --json | Output machine-readable JSON |

tensorcost estimate \
  --prompt "Summarize the following document in three bullet points." \
  --model anthropic/claude-3-haiku \
  --calls-per-day 5000

Compare every Anthropic model (direct API + Bedrock + Vertex + OpenRouter) for the same prompt:

tensorcost estimate \
  --provider anthropic \
  --tier latest \
  --prompt "Hello world" \
  --calls-per-day 5000

List all Bedrock-hosted models:

tensorcost models --provider bedrock

Deprecated models (claude-2, claude-instant) are hidden by default; pass --include-legacy to show them.

tensorcost models lists list prices in dollars per million tokens ($/M), not per-call cost. Use estimate to see what a specific prompt actually costs. The live catalogue is refreshed daily from TensorCost pricing plus LiteLLM gap-fill (~2,500+ models including Bedrock, Azure, Vertex, Groq, etc.).

compare

Compare cost between a baseline and a candidate prompt.

| Flag | Description | |------|-------------| | --baseline <text> / --baseline-file <path> | Baseline prompt | | --candidate <text> / --candidate-file <path> | Candidate prompt | | --model <id> | Model identifier (required) | | --expected-output-tokens <n> | Override output token estimate for both sides | | --json | Output machine-readable JSON |

tensorcost compare \
  --baseline-file prompts/v1.txt \
  --candidate-file prompts/v2.txt \
  --model openai/gpt-4o \
  --json

models

List models from the live catalogue (or the offline seed) and their input/output rates.

| Flag | Description | |------|-------------| | --provider <name> | Filter by provider (openai, anthropic, bedrock, azure, vertex, …) | | --json | Output the full catalogue as JSON |

tensorcost models
tensorcost models --provider openai
tensorcost models --json

Model catalogue

The CLI does not hardcode the full provider catalogue. On each run it:

1. Starts from a small bundled seed (models.json) so offline / CI works with zero network.
2. Merges the daily CloudFront snapshot at https://console.tensorcost.com/cli/model-rates.json (TensorCost ai.model_pricing + LiteLLM gap-fill), cached for 1 hour at ~/.tensorcost/rates-cache.json.

Use full ids (openai/gpt-4o). Bare slugs (gpt-4o) resolve when unique; if several providers match, the CLI lists the candidates and exits.

Run tensorcost models (or tensorcost models --provider anthropic) to see what is currently known.

Environment variables

| Variable | Description | |----------|-------------| | TENSORCOST_CATALOG_URL | Override the catalogue URL. Defaults to https://console.tensorcost.com/cli/model-rates.json. Point at staging with https://console.<env>.tensorcost.com/cli/model-rates.json. | | TENSORCOST_OFFLINE | Set to 1 / true to skip the network fetch and use the bundled seed (plus any on-disk cache). |

Live rates

No API key is required for rates — the snapshot is public list pricing. Any model in the live response takes precedence over the bundled seed. On network / HTML / schema errors the CLI proceeds silently with the bundled table (and a stale cache when present).

Tokenization

Token counts are estimated using ceil(charCount / 4), which is accurate to within ~10% for typical English text. If tiktoken or @anthropic-ai/tokenizer is installed in the same environment, the CLI will use it automatically — no configuration needed.

JSON output

Both estimate and compare accept --json for CI / scripting:

result=$(tensorcost estimate --prompt-file prompt.txt --model openai/gpt-4o --json)
total=$(echo "$result" | jq '.totalCostUsd')

PR review workflow

# Makefile target — compare your branch's prompt against main
.PHONY: cost-check
cost-check:
	tensorcost compare \
	  --baseline-file $(shell git show main:$(PROMPT)) \
	  --candidate-file $(PROMPT) \
	  --model $(MODEL)