Coinbase launches usage‑based agent pricing
Coinbase’s developer platform rolled out an 'Upto' usage pricing option that ties fees to agent compute, covering inference and related data costs for agent workloads. The change reflects a broader industry move to align billing with variable agent execution rather than flat token pricing. For teams building agentic flows, it’s an early example of vendor pricing tailored to long‑running, tool‑heavy workloads. (x.com)
Coinbase just changed a small but painful part of building artificial intelligence agents: the bill no longer has to be a single fixed number before the work starts. Its new “Upto” option lets a seller set a ceiling and then charge only for the compute and data the job actually used. (cointelegraph.com) That fixes a mismatch in how many agents work. A simple request might need one model call, while a harder request can trigger extra reasoning, more web lookups, and longer runtime, so the real cost is not known at the first second. (docs.x.ai) Most artificial intelligence pricing still starts with tokens, which are the chunks of text a model reads and writes. OpenAI’s pricing page, for example, lists model costs per 1 million input and output tokens, and then adds separate charges for tools like web search, file search, and code execution. (developers.openai.com) Other providers already price tool-heavy requests this way. xAI says requests that use server-side tools are billed from two parts at once, token usage and tool invocations, because an agent can decide on its own how many tools to call. (docs.x.ai) Coinbase’s move sits inside x402, a payment protocol it originally built for machine-to-machine payments on the web. The protocol revives the old Hypertext Transfer Protocol 402 “Payment Required” status code so a client can discover a price, pay programmatically, and retry the request without opening an account or session. (docs.cdp.coinbase.com) Until now, x402 was better suited to deterministic services where every call costs the same amount. Coinbase said the old model worked for exact fixed-price payments, but blocked services where cost changes with token count, compute time, or query complexity. (cointelegraph.com) The new setup is closer to how a gas pump preauthorization works. The buyer approves a maximum amount up front, and the server settles only the amount needed to finish the task, which Coinbase says can leave the buyer paying less than the cap. (cointelegraph.com) Coinbase also says this version is built for Ethereum Virtual Machine networks, supports all ERC-20 tokens, and can use the Coinbase Developer Platform facilitator for gasless payments. In x402, a facilitator is the service that verifies payment payloads and settles them onchain for the server. (cointelegraph.com) (docs.cdp.coinbase.com) This is not Coinbase’s first usage-based meter. Its Node application programming interface already bills by “Billing Units,” where a basic method can cost 30 units, a heavier method 100 units, and advanced tracing calls 500 units, with usage above the free tier priced at $0.50 per million units. (docs.cdp.coinbase.com) The timing is also notable because Coinbase moved x402 into a neutral home on April 2, 2026. The Linux Foundation says Coinbase contributed the project to the new x402 Foundation, which is meant to govern the protocol as an open standard for autonomous and machine-to-machine payments. (linuxfoundation.org) So the story here is less “Coinbase changed a pricing page” than “agent bills are starting to look like cloud bills.” When the work includes model inference, tool calls, and data fetches that expand and shrink mid-run, a flat fee stops fitting the job. (developers.openai.com) (docs.x.ai)