Humanbound website
Firewall OPEX Estimator

Runtime protection that learns from your tests.

The Humanbound Firewall is a four-tier defense layer between your users and your AI agent. Tiers 0–2 run locally with zero token cost. Tier 3 (LLM Judge) handles edge cases. Estimate your monthly cost below.

Four-Tier Architecture

T0Sanitization — Strips invisible chars, bidi overrides.Free
T1Basic Attack Detection — DeBERTa, ensemble.Free
T2Agent-Specific Classifier — Trained on your test data.Free
T3LLM Judge — Deep contextual analysis, only when uncertain.Tokens
5,000
50025K50K75K100K
12%
5% (well-trained)15%30% (no training)
Your agent runs on

Used for comparison only. Does not affect firewall cost.

Firewall judge model

A cheaper model works well here. Classification task, not creative generation.

Estimated Monthly Firewall OPEX

Tier 3 Calls

600

per month

Monthly Cost

€0.13

judge model tokens

Per Conversation

€0.0000

all tiers combined

Cost Breakdown

Tiers 0+1+2 (local models)
€0.0000 (0%)
Tier 3 input tokens
€0.07 (57%)
Tier 3 output tokens
€0.05 (43%)

Firewall Cost vs Agent LLM Cost

Agent LLM cost (estimated)€37.50
Firewall cost (Tier 3 only)€0.13

Firewall adds ~0.3% to the agent's LLM cost.

Assumptions

  • Tiers 0, 1, 2 run as local models on your compute. No token cost. Negligible CPU/GPU footprint.
  • Tier 3 (LLM Judge) is called only when Tiers 1-2 cannot make a confident decision. Default escalation: ~10-15%. Better Tier 2 training reduces this.
  • Tier 3 token estimate per call: ~800 input tokens (system prompt + policy + user message) and ~150 output tokens (verdict + explanation).
  • Agent LLM cost assumes ~1,000 input + 500 output tokens per conversation at the agent model's pricing (for comparison only).
  • The firewall judge and your agent can use different models. A cheaper model often works well for the judge since it's a classification task.
  • Prices are fetched live from the LiteLLM community model database. If the fetch fails, defaults from May 2026 are used.
  • These are estimates. Actual costs depend on conversation length, policy complexity, and Tier 2 training quality.