A cōdel is a model implemented as a blend of deterministic code and AI. The split is a choice, not a given: some cōdels are 90% code and 10% AI, others are 50/50, others lean the other way. The interface is always the same — openrouter-compatible, drop-in for any existing client.
Where a foundation model predicts every token, a cōdel computes what it can and spends AI only where computation isn't enough.
There's no fixed recipe. The ratio reflects what a domain requires: exact, auditable logic becomes code; ambiguous, open-ended judgment stays on AI — until it doesn't need to.
Every cōdel starts as whatever mix is practical and evolves. As behavior is exercised and proven, the AI components that are unnecessary or risky are progressively eliminated and replaced with code.
AI-heavy → hybrid → lean → pure code
The result: models that get faster, cheaper, more deterministic, and more auditable over time — while the risk surface of unpredictable output shrinks with every release.
Most cōdels are AI-model agnostic. The AI component isn't welded to a specific foundation model — it's a pluggable slot that can target any model on openrouter.
The code is fixed. The AI is your choice.
opencodel.com runs an openrouter-compatible passthrough proxy. Point your client at a single endpoint and it routes both kinds of traffic:
One key. One endpoint. One billing surface. The client never needs to know which path a request took — the response is identical.
For direct openrouter calls, your existing keys work as-is — bring your own key, keep your existing billing, and the proxy forwards your traffic with full drop-in compatibility. No re-provisioning, no new accounts, no client changes.
Pure models charge for effort — tokens in, tokens out. cōdels charge for outcome: what was actually accomplished. A cōdel that resolves a request in 30 lines of code instead of 2,000 tokens is priced on the result, not the machinery.
And because a cōdel's behavior is provable — deterministic, tested, audited — every charge can be verified against what it did.
If a cōdel doesn't deliver the promised outcome, you get your money back. Full refund, no quota of exceptions, no dispute theater. The same audit trail that makes cōdels trustworthy to run makes refunds trustworthy to grant.
When was the last time Anthropic refunded a hallucinated inference that broke your repo? Exactly.
opencodel.com is where cōdels are published, benchmarked, and traded. Each listing ships with its spec, its test suite, and benchmarks against the foundation models it replaces — including its current code/AI ratio and its evolution history.
Listings are coming online. Endpoint: https://opencodel.com/v1/chat/completions
cōdels aren't confined to chat responses. Their code can act through trusted, verified interfaces — the same channels people actually use — with every action logged and auditable:
The same audit trail that governs the code governs its side effects.
Behind the marketplace is the Coddler — the foundry that makes creating cōdels fast. Design, test, and deploy a new cōdel in hours, not weeks — because cōdels care, literally:
Cōdels care. Literally.
Have a workflow that needs a cōdel built to fit? We run a forward-deployed engineering service — our engineers embed with your team, design the cōdel against your actual workloads, and ship it into your stack.
You pay for outcomes in production — not for the build.