Electricity wasn’t useful until it was metered. The internet wasn’t real until packets had a cost.
AI compute is next, and today, it’s still priced like a mystery.
Intelligence becomes infrastructure.
OpenAI prices in one unit, Anthropic in another, Google in a third, every other provider in its own. No conversion. No benchmark. No way to know if $0.015 per 1K tokens is cheaper than $0.003 per 1K. The CPI normalizes every model in the basket into one equal-weighted reference, the SCU.
From black box to open signal.
When the cost of compute is published on-chain and updated when the basket is reconstituted, budgets become precise and capital flows where it should. Builders price their inference against a versioned, verifiable benchmark, not a provider’s docs page from last Tuesday.
A planet that runs on priced thought.
A startup in Lagos and a lab in San Francisco query the same oracle, get the same SCU rate, build against the same standard. Intelligence becomes a market priced by one shared reference — not pricing pages each provider updates on its own clock.
The standard hasn’t been set.
No equivalent exists. AI inference pricing changes weekly and no one tracks it programmatically. The CPI spans 9 providers and 19 models, each weighted equally with no tiers. Reconstitution-driven and versioned on-chain. The public API is open to all.
What a priced world unlocks.
Once compute has a verifiable reference price, the things that depend on one become possible.
Programmatic price discovery
Query current per-model pricing and the aggregate SCU through one public API. Versioned, free reads, no authentication. Your routing logic, your decision, built on a benchmark that doesn’t shift when a provider rewrites its pricing page.
A benchmark to audit against
The SCU is the cost of a standardized workload, 1,000 input plus 500 output tokens, taken as the equal-weighted geometric mean across the basket. Convert your actual usage into the same units and you have a public reference point to measure AI spend against the market rate.
Composable compute cost
The CPI is published via the OracleRegistry contract on Base. Smart contracts read the current SCU directly — the same way they read ETH/USD from Chainlink. AI cost data becomes a primitive other on-chain systems can compose against natively.
For AI agents — Model Context Protocol
The official Compute Finance MCP server gives any LLM client read-only access to the oracle, cost estimation, and session cost reports. Stdio transport. No API key required.
data
compute
render
analyze
history
Install in any MCP client
npx @compute-finance/mcpClaude Code (registers MCP + skills + cost hook)
npx @compute-finance/mcp setupReference:github.com/compute-finance/mcp/.well-known/mcp/server-card.json