See more like this on Google
Share article
Share article
The irony is hard to miss: crypto keeps promising machine-to-machine commerce, then charges like a vending machine that only takes exact change. Coinbase says x402 is fixing that, adding usage-based pricing for AI compute so services can bill by what was actually consumed instead of forcing a flat fee upfront. [1]
The update, announced Thursday by Coinbase Developer Platform, introduces what it calls the "Upto" scheme. The point is simple enough. x402 previously handled exact, fixed-price payments well, which works for deterministic APIs where every request costs the same. It worked far less well for AI workloads, where cost can swing based on token count, compute time, or query complexity. LLM inference is not a soda can. Everyone definitely predicted that billing nuance would matter. [2]
Enjoy articles without ads?
Register for free and get unlimited access to all articles.
What changed in x402
Under the new model, developers can authorize payments up to a defined amount for a request, rather than hard-coding one exact price before the job starts. That opens the door for metered AI services, including language model calls, compute-heavy tasks, and data queries that do not have a clean fixed cost at the outset.
Coinbase said the implementation is Ethereum$1,686.33 Virtual Machine based, supports all ERC-20 tokens, and can be used with its CDP Facilitator for gasless payments. That last part matters more than the marketing gloss suggests. If every micro-transaction needed separate gas management, the whole "internet-native payments for agents" pitch would collapse under its own friction. [1]
Why fixed fees were a bottleneck
Flat pricing is fine when an API returns the same kind of response every time. AI services usually do not. One prompt can be cheap, another can eat far more tokens, latency, and GPU time. If billing cannot adapt, providers either overcharge to cover worst-case usage or undercharge and absorb the difference. Neither is a great recipe for automated commerce.
That gap is one reason x402 has been framed as infrastructure for AI agents, software that can discover services, request work, and pay programmatically. Metered billing makes that story at least more plausible. Not automatically good, just less obviously broken. [3]
Why this matters for agentic AI
The broader thesis around x402 is that AI agents need payment rails built for APIs, not humans clicking checkout buttons. Usage-based pricing gets closer to how cloud and AI services are actually consumed, especially for inference and compute. It also creates a cleaner path for providers selling granular access to models, datasets, and processing power.
Research around x402 has consistently centered on this "agentic commerce" angle: autonomous software buying services on demand. The protocol's appeal is not just crypto settlement for its own sake, but the ability to attach programmable payments directly to HTTP-style service requests. If that works reliably, AI tools can pay for what they use at the time they use it. [4]
The real test
Of course, adding metered billing is easier than proving anyone wants it at scale. The hard part is adoption by API providers, agent frameworks, and infrastructure marketplaces. Developers also need safeguards around spending caps, failed requests, disputed usage, and billing transparency. "Up to" pricing sounds elegant until someone discovers their autonomous bot has expensive taste. [5]
Looking Ahead
x402 now looks more aligned with real AI economics than it did under fixed-fee billing. That is the practical takeaway. The protocol can finally handle services whose costs vary by usage, which is common across LLM inference, compute jobs, and complex data queries.
What to watch next is not the announcement, but implementation. Are AI platforms integrating it? Are providers exposing metered endpoints through x402? And does gasless ERC-20 billing make autonomous purchasing meaningfully smoother, or just more cryptographically decorated? That answer will decide whether this is useful infrastructure or another tidy demo for a market still waiting on demand.
