Early access — cascade metrics are real (derived from canonical token telemetry); the operator field is a curated seed. Learn more about the data
\u25C8 SigRank vs CostHawk

Consumption Counts. Efficiency Matters.

CostHawk has a privacy-first anonymized leaderboard ranked by total token consumption. SigRank ranks by Yield efficiency. Both are privacy-preserving. Only one measures skill.

The short version: CostHawk

CostHawk (costhawk.ai) is a privacy-first AI tools leaderboard. It tracks Claude Code, Codex, and Cursor, anonymizes operator data, and ranks by total token consumption. For seeing how much you consumed compared to others in a privacy-respecting way, it does the job. But it ranks by volume: who burned the most tokens. SigRank ranks by efficiency: who compounded their tokens the best.

The analogy: CostHawk is the odometer. It tells you how far you drove. SigRank is the MPG readout. It tells you how efficiently you drove. Both are privacy-preserving. Only one tells you if you are a good driver.

Feature comparison

FeatureCostHawkSigRank
What it measuresTotal token consumption across AI toolsToken cascade efficiency (Yield, Leverage, SNR, Velocity)
Headline metricTotal tokens consumedYield (Υ) = (cache_read × output) / input²
What it tells youHow much you consumed: volume onlyHow efficiently you use AI: are tokens compounding?
Leaderboard rankingBy total token consumption (anonymized)By Yield efficiency (signed, verifiable)
Tools trackedClaude Code, Codex, Cursor19+ AI coding agents with dedicated adapters
Cascade efficiency score (Yield)NoYes
Class tier (IGNITER to ARCH+)NoYes
Build archetype (10 types)NoYes
Operator profiles + head-to-head compareNoYes
MCP server for AI-agent integrationNoYes
ed25519-signed snapshot submissionNoYes
Privacy-preserving (token counts only)Yes (anonymized)Yes (token counts only, never prompt content)

Why consumption is not efficiency

CostHawk answers "how much did I consume compared to others?" That is useful for benchmarking spend. But it does not tell you if you are good at using AI. Two operators can consume the same total tokens and get wildly different outcomes. One reuses cached context efficiently and produces 30K output tokens. The other re-sends the same context every turn and produces 3K. Same consumption, ten-fold difference in signal.

SigRank's headline metric, \u03A5 = cache_read \u00D7 output / input\u00B2, measures exactly that gap. It rewards the operator who compounds cached context into output and penalizes the one who burns fresh input without leverage. CostHawk gives you the total. SigRank tells you whether the cascade that produced that total is compounding or burning.

The four token pillars (SigRank reads these)

  • Input: tokens you send to the model
  • Output: tokens the model generates back
  • Cache-read: cached tokens reused from prior context
  • Cache-write: new tokens written to cache for future reuse

See your efficiency score

If you already use CostHawk, you have the raw token counts. SigRank reads the same logs and adds the efficiency layer:

npx sigrank          # read your logs, show your cascade
npx sigrank me       # see your yield, archetype, and class tier
npx sigrank submit   # sign + publish to leaderboard

Prefer to inspect before you submit? Run sigrank me --dry-run to see your scored payload locally, or paste your token counts into the /score calculator for an instant read.

Frequently asked questions

Is SigRank a CostHawk alternative?
They overlap on privacy positioning but measure different things. CostHawk has an anonymized AI tools leaderboard that ranks by total token consumption. SigRank ranks by Yield efficiency. CostHawk tells you who consumed the most tokens. SigRank tells you who used tokens the best. If you want a privacy-first consumption leaderboard, CostHawk does that. If you want a privacy-first efficiency leaderboard, SigRank is the only one.
What does CostHawk not measure that SigRank does?
CostHawk reports total token consumption. SigRank derives the cascade architecture from token counts: Yield (cache_read times output divided by input squared), compression ratio, SNR, Leverage, and Velocity. CostHawk sees the total. SigRank sees the structure. Two operators can consume the same total tokens with wildly different efficiency. One reuses cache and produces high output. The other burns fresh input and produces little. CostHawk ranks them the same. SigRank ranks them 10x apart.
Can I use both CostHawk and SigRank?
Yes. CostHawk is a consumption leaderboard; SigRank is an efficiency leaderboard. They read from similar local session logs and do not conflict. Use CostHawk for anonymized consumption tracking across Claude Code, Codex, and Cursor. Use SigRank for efficiency scoring, build archetype classification, class tier assignment, and leaderboard ranking by Yield. Many operators use a consumption tracker for cost monitoring and SigRank for skill measurement.
Which is better for privacy-conscious operators?
Both are privacy-preserving. CostHawk anonymizes its leaderboard. SigRank uses ed25519-signed submissions with token counts only: no prompts, no code, no transcripts. The difference is what they do with the privacy guarantee. CostHawk uses it to rank consumption anonymously. SigRank uses it to rank efficiency verifiably. Privacy without measurement is just anonymity. Privacy with efficiency measurement is a verifiable skill ranking.
Why does CostHawk only track three tools?
CostHawk currently tracks Claude Code, Codex, and Cursor. SigRank supports 19+ AI coding agents with dedicated adapters, including Claude Code, Codex CLI, Gemini CLI, Copilot CLI, Amp, Qwen Code, Goose, OpenCode, Kilo CLI, Hermes Agent, Devin, OMP, Pi, OpenClaw, Droid, and Codebuff. If you use tools beyond the big three, SigRank has wider coverage. See /platforms for the full list.

Ready to see your cascade?

Keep your CostHawk consumption tracking. Add the scoring, the leaderboard, and the operator profile that turns those readings into a rank. Install SigRank and submit your first signed snapshot in under a minute.

Related: vs ccusage · vs Tokscale · vs clawdboard · Supported Platforms