1. Inference costs vary wildly across providers and models. Claude Opus 4.1 is ~$75 per million output tokens, while Gemini 2.5 Pro is ~$15. That volatility makes budgeting hard, especially if you want portability across labs.
2. The entire economy of API credits annoys me to no end.
3. More blue sky: I've been experimenting with agents with financial self-sovereignty. It seemed strange to me that there's no way an agent can call for more or different inference without having to be pre-seeded with some credential. I wanted long-context, long-term deployments.
With all of that in mind, we’ve been experimenting with Melange (MEL): a unit pegged to the most expensive frontier model’s output cost (currently $75/MTok for Claude Opus). By definition, 1 MEL should always buy at least 1M output tokens, no matter which provider you’re using.
This is meant to work a bit like a stable reference for compute: instead of juggling provider-specific credits, you hold MEL and redeem it for inference anywhere. At the moment, it runs through our infra (until we partner with more providers), but we’re testing on-chain exchange where you can swap MEL directly for inference using only your wallet keys.
This is done using a smart contract that escrows your MEL for a duration you configure. You then simply sign your inference requests with the same key that holds the MEL to have them processed. When the deadline hits, the remaining MEL flows back to your account. This detail is abstracted away using a wrapper for the OpenAI library that handles the signing.
The smart contracts are open source at the moment; we’ll be opening all mission-critical parts of the stack as well eventually.
Paralus•4mo ago
1. Inference costs vary wildly across providers and models. Claude Opus 4.1 is ~$75 per million output tokens, while Gemini 2.5 Pro is ~$15. That volatility makes budgeting hard, especially if you want portability across labs.
2. The entire economy of API credits annoys me to no end.
3. More blue sky: I've been experimenting with agents with financial self-sovereignty. It seemed strange to me that there's no way an agent can call for more or different inference without having to be pre-seeded with some credential. I wanted long-context, long-term deployments.
With all of that in mind, we’ve been experimenting with Melange (MEL): a unit pegged to the most expensive frontier model’s output cost (currently $75/MTok for Claude Opus). By definition, 1 MEL should always buy at least 1M output tokens, no matter which provider you’re using.
This is meant to work a bit like a stable reference for compute: instead of juggling provider-specific credits, you hold MEL and redeem it for inference anywhere. At the moment, it runs through our infra (until we partner with more providers), but we’re testing on-chain exchange where you can swap MEL directly for inference using only your wallet keys.
This is done using a smart contract that escrows your MEL for a duration you configure. You then simply sign your inference requests with the same key that holds the MEL to have them processed. When the deadline hits, the remaining MEL flows back to your account. This detail is abstracted away using a wrapper for the OpenAI library that handles the signing.
The smart contracts are open source at the moment; we’ll be opening all mission-critical parts of the stack as well eventually.