AI model replacement, measured
Find the cheapest AI model that still passes your real work.
Upload real tasks from your Claude, GPT or Gemini workload. Harpd tests cheaper models against them and measures quality, cost, latency and reliability before you switch.
Your workload, not a leaderboardQuality + cost, measuredSwitch only when safe
What Harpd does
Harpd helps you measure what your AI workload actually costs, optimize it by finding cheaper models that pass your real tasks, and control production spend with hard limits — so you spend less and never overspend.
Pick the problem that sounds like yours.
My AI bill is too high
See what every major model would cost for your exact workload, ranked cheapest first.
AI Cost Calculator →I don't know which cheaper model is safe
Test candidate models against your real tasks and switch only when the evidence holds.
ModelSwitch →My AI agent can overspend
Put hard budgets, approvals and reconciliation around production agent payments.
Spend Control →Autonomous agents spend real money — often with no one watching.
A single misconfigured agent can issue thousands of paid API calls without a human in the loop. Spend Control makes that spend visible and bounded before it becomes a bill. Read how on the blog →
Measure
Know what your AI workload actually costs — by model, agent, and task.
Optimize
Find cheaper models that pass your real tasks, with evidence, not a public leaderboard.
Control
Stop production agents from overspending with budgets, approvals, and reconciliation.
Measure before you switch.
All free, no signup, runs in your browser.
AI Cost Calculator
Estimate your monthly bill, ranked by model.
LLM Pricing
Side-by-side list prices for 13 hosted models.
Compare AI Models
Two models, one workload, cleanest price gap.
Cost per Successful Task
The metric production systems actually pay against.
AI Agent Cost Calculator
For autonomous agents making thousands of paid calls.
LLM Cost Calculator
Model-by-model token math, no signup.
Open Benchmarks
Public CSV + JSON on GitHub. Verify before you trust.
AI cost intelligence guides.
Aug 19, 2026
Debunking the Myth: Why Cheap-Per-Token AI Models Can Inflate Overall Costs
"Debunking the myth of cheap-per-token AI models & their hidden costs. Learn how optimizing for Cost per Successful Task saves up to 90% (McKinsey)."
Read postAug 18, 2026
Cost-Effective LLMs for JSON Extraction, Classification, and Code Review: A Comparative Analysis
Compare cost-effective LLMs for JSON extraction, classification, and code review. Analyze top options, capabilities, and integrate with observability tools like @harpd/observe.
Read postAug 15, 2026
How to Audit AI Agent Transactions: A Practical Guide to Schema, Spend, and Compliance
Learn how to audit AI agent transactions with a shared audit schema, track spend, and ensure compliance in finance, healthcare, and support use cases.
Read postOpen-source SDKs for agent observability, audit, and budget governance.
@harpd/observe
Zero-dependency transaction observability for AI agents.
agent-budget-policy
Declarative budgets and approval rules for agent payments.
x402-logging-middleware
Request/response logging for x402 payment flows.
agent-transaction-audit-schema
A portable schema for agent transaction audit trails.
mcp-paid-tool-starter
Starter kit for paid MCP tools and per-call billing.
Experiments live under Harpd Labs.
Stillwell (sleep), CrewProof (cleaning), ImgKit (gifts) and Fate (games) are separate experiments on their own subdomains — kept off harpd.com so this site stays focused on AI cost intelligence.
Harpd Spend Control, in plain terms
What does Harpd do?
Harpd is AI Cost Intelligence. We help teams measure what their AI workloads actually cost, find cheaper models that can safely replace expensive ones, and put hard spending limits around production AI agents — so you spend less and never overspend.
What is an AI agent payment?
An AI agent payment is a machine-initiated transaction where an autonomous agent pays for an API, tool or service on your behalf — using protocols like x402, MPP or AP2. Spend Control attributes and gates those payments.
How does Spend Control stop agents from overspending?
It intercepts paid calls before settlement. You set per-agent budgets, allowlists and an approval threshold; calls that exceed a limit are blocked or routed to a human for review, then every payment is reconciled against the policy afterwards.
Does Spend Control hold my money or private keys?
No. Spend Control is non-custodial. It observes and gates spend through your existing payment stack — it never takes custody of funds or signing keys, so there is no new counterparty risk.
Which AI agent payment protocols does it support?
Spend Control works across the three leading agent-payment protocols: x402, MPP and AP2. The same policy layer applies regardless of which protocol an agent uses to pay.
What is the difference between Observe and Enforce mode?
Observe mode attributes and reports spend with no blocking — safe to switch on immediately. Enforce mode applies budgets and approvals before settlement, actively preventing overspend. You can start in Observe and move to Enforce when confident.
Are Stillwell, CrewProof, ImgKit and Fate part of Spend Control?
No. They are separate experiments built under Harpd Labs: Stillwell (sleep and calm), CrewProof (cleaning crew scheduling), ImgKit Gift Studio (personalized digital gifts), and Fate (a symbolic life roguelike). They live on their own subdomains and a single Harpd Labs link — Harpd.com stays focused on AI cost intelligence.
Get started
Test 20 tasks free