InferHub

InferHub is a per-token marketplace and reverse proxy for AI inference that pools upstream LLM subscriptions behind a single OpenAI- and Anthropic-compatible API. It allows users to deposit USDC and pay only for the tokens they actually use, eliminating the need for monthly subscriptions. The platform aggregates access to models like Claude, Codex, Gemini, and GLM through one endpoint.

Consumers can attach a bid to their request, expressed as a multiplier of their maximum input and output price, to route to a provider whose ask meets their cap. If no bid is provided, the system picks the best-scoring provider regardless of price. Every provider ask is hard-capped at 50% of the official pricing, ensuring users never pay more than half the upstream's published rate.

Publishers set their own per-token rates for the models they connect, earning a margin on every successful request routed through their subscriptions. Earnings accumulate in USDC and can be transferred to a spendable balance or withdrawn to a wallet via a payout request after verification and reconciliation.

The marketplace currently supports 350 models across text, image, and video categories, with 575 active providers and 27 curated upstreams. Prices fluctuate in real-time, and the system automatically routes auto-routed models to the cheapest available provider across all marketplaces without requiring a provider prefix.

Balances added to an account are non-refundable because funds are consumed in real time as compute is delivered. Charges for successfully completed requests are final, and unused balances remain available for future requests but are not refunded as cash unless required by law or in cases of duplicate deposits.

Users can add funds using USDC on Solana mainnet, other supported cryptocurrencies, or third-party processors like PayPal, debit cards, credit cards, QRIS, and Binance Pay. Confirmed crypto deposits are typically credited within about a minute, though the platform is not responsible for losses from on-chain errors or wallet misconfiguration.

Failed requests, requests with no output, and free models cost $0, while every billed request costs at least $0.00001. If a request's token count results in a charge lower than this minimum, it is rounded up to $0.00001 to ensure accurate billing.

The platform operates out of Indonesia, with a physical address at Jl. Siwalankerto Permai II No. C35 Wonocolo, Surabaya 60236. Support is available via email at [email protected], phone at 089677717105, or through a community Discord server for quick questions.

Some users on Reddit question why the pricing appears so cheap compared to original models, noting that the platform pools subscriptions to offer competitive rates. The live market pricing page displays current rates for various models, which can vary significantly based on upstream availability and demand.

The system waits up to 30 seconds by default, configurable up to 120 seconds, for a free slot if a user sets a maximum desired rate in a Budget. If no slot opens within the timeout, the system returns an error instead of overcharging the user.

A first-time customer should know that all usage, billing, and prices are recorded and charged in USD, with Rupiah figures provided only as approximate conversions for reference. The platform supports prompt content checks before forwarding requests to ensure safety and compliance.

The API is compatible with both OpenAI and Anthropic standards, allowing developers to integrate with existing tools. Publishers must comply with the upstream provider's terms and represent that they own or are authorized to use the subscription they are routing.

The website states that the Operator is an individual developer, and the service supplements the Terms of Service and Privacy Policy. The platform claims to be built for the open AI ecosystem, facilitating a decentralized marketplace for inference.

Some users on Reddit have asked if the model is legal, questioning how providers can sell AI models so cheaply. The platform addresses this by aggregating multiple upstream subscriptions to create a shared pool that lowers costs for consumers.

The official website of InferHub is inferhub.dev.