The guard is billed by what it actually consumes: input tokens. It generates no text, so there is no output charge — ever. Prepaid credits; when they run out the guard fails closed (answers "cannot foresee", which every tool treats as risky) rather than silently going away.
A check is ~1.5k input tokens → ≈ $0.00006 per check ($0.06 per 1,000 checks). A $5 credit pack buys 125M tokens ≈ 83,000 checks. Pay in USDC or USDT on Polygon, Arbitrum, Base or Ethereum. At typical API inference prices ($0.10–0.60/1M) the guard adds a small fraction to the bill; if you self-host your inference, the self-hosted guard costs zero.
$0.04 per 1M input tokens
No output charges: a check is ~1.5k input tokens and the model generates none — the answer is read from the logits. Prepaid credits, no subscription, no minimum. Out of credits the guard fails closed (cannot foresee → treated as risky): it never silently turns off.
≈ $0.00006 per check · $0.06 per 1,000 checks
Free forever, Apache-2.0
Open weights on Hugging Face: 27B (bf16, FP8, INT4) and the MLX 4-bit build for Apple Silicon. Same API, same prompt format, same calibration. Mac, laptop or your own GPUs.
For inference you self-host, the guard can run beside it at zero marginal cost (measured on shared GPUs).
from $5k setup + retainer
We install the guard next to your production LLM — measured on 4× B200 sharing GPUs with a production SGLang deployment, isolated venv, monitored before/after. Runbook, launcher, monitoring and SLA.
Retainer from $1k/mo
from $10k your stack as the world
We generate consequence data for YOUR domain — your infra, your billing, your compliance rules — and train the model on it. You keep the weights.
The highest-margin option; the model is the proof of concept.
Every claim links to the measurement it comes from. Re-run the suites yourself against your own deployment — the data ships with the cookbook.
Agent harm with the guard consulted first
48% → 6% (−41.4 pp, 95% CI [26.2, 56.7], n=93, one seed)
full studyClassic destructive git commands flagged
8/8 at 98–99% · 0 false alarms on 9 safe commands
guard batteryNever used a consequence model? Read the getting-started guide first: how to write a state, the question types and how to read the probabilities, with real answers.
pip install ekbasis
export EKBASIS_URL=https://openinterp.org/api/v1 # or your self-hosted server
export EKBASIS_API_KEY=ekb_... # /console → keys
cd your-repo && ekbasis git-check -- "git reset --hard"
# Ekbasis: RISKY (lose uncommitted work: 99%) → exit 2Questions and enterprise: [email protected] · Papers: Look When Unsure, Check When Sure · When Does a Consequence Model Make AI Agents Safer?
Open the console: keys, credits and usage