x402 Ditches Flat Fees for 'Upto' Usage-Based AI Compute Pricing

Coinbase has finally stopped treating AI agents like they're ordering off a value menu. The x402 protocol just got a glow-up with usage-based pricing for compute requests, replacing the old flat fee model that was about as nuanced as a sledgehammer.

The new 'Upto' scheme went live this week, enabling variable-cost services for agentic AI like LLM inference, compute and data queries. Previously, x402 only supported exact fixed-price payments, which worked for deterministic APIs but blocked services where costs depend on usage, token count, compute time, or query complexity. Basically, it was trying to fit square pegs into round holes while the AI revolution happened in real-time.

Upto is an EVM implementation supporting all ERC20s, with CDP Facilitator enabling fully gasless payments. For those keeping score at home, this means you can pay for your AI's thinking time without actually thinking about gas fees. Revolutionary stuff.

Under the new system, sellers configure maximum prices while buyers authorize spending up to a specific amount. Servers charge only what it actually costs to complete the task, so users won't overpay and may pay less than the maximum. It's like tip jars, but for compute—and the AI actually earns it.

Previously, simple and complex AI agent requests cost the same, leaving users either overpaying or underpaying. Now users can set their price ceiling before a task instead of guessing. Imagine paying $5 for a cheeseburger and $5 for a wagyu steak—sure, technically fair, but your wallet had opinions.

Developed by Coinbase, protocol ownership was transferred to the Linux Foundation earlier this month. Google, Microsoft and Amazon Web Services hold stakes via the x402 Foundation. Big Tech is in the building, folks, and they're not just here for the snacks.

Despite the hype, x402 adoption has declined in 2026 after hitting 13.7 million transactions during the week of November 4-10 (its biggest week). Weekly volume dropped below 1 million in early January and fell to just 112,708 transactions by late March. Nothing says "promising technology" like a chart that looks like a bungee jump gone wrong.