Pay-Per-Call x402 vs. API Keys in Brave's Agentic Search Proxy

When wiring web search into an agent harness, you now face a concrete billing choice on Brave Search. You can provision a standard API key backed by a prepaid credit balance, or point your agent at Brave's new agentic payments proxy (search.agent.s.brave.app), launched on November 12, 2026, where every request settles directly in USDC on Base (eip155:8453) at $0.005 per call via an x402 V2 client such as Stripe's Purl CLI wallet.

Skipping account signup and secret rotation makes ephemeral CLI agents faster to ship. Brave also had a direct supply-side motive. On November 11, Brave detailed its anti-scraping rollout, noting that Google's google.com/goto link redirects had shifted automated search scrapers onto Brave's index at spikes of several thousand requests per second. With Microsoft having retired the Bing Search APIs and Google's Custom Search JSON API shutting down on January 1, 2027 according to Brave's API portal, putting an open 402 Payment Required meter at $5 per 1,000 queries turns unmonetized crawler load into revenue while opening a keyless path for legitimate tools.

Before switching an agent pipeline over to per-request crypto payments, read Brave's proxy specification (llms.txt). Its list of excluded endpoints shows exactly where gross per-call settlement breaks down compared to a standard prepaid API ledger.

A traditional API account acts as an internal credit ledger, the pattern economist Perry Mehrling (The Economics of Money and Banking, Lecture 2) separates from final settlement. Your service funds a balance once, fires thousands of queries of varying cost, and nets the total through a periodic card charge or invoice. Brave's search.agent.s.brave.app proxy removes that netting layer completely: no session, no credit balance, and no top-up. Each call must carry its own payment credential and clear individually through Brave's facilitator.

That single architectural shift creates three practical boundaries for agent builders.

First, sub-cent endpoints do not clear one request at a time. Brave left Autosuggest and Spellcheck off the x402 proxy because, as its documentation states, "they cost less than the fee a facilitator takes for settling one payment, so they wait on settling many queries together as one payment." On an internal prepaid ledger, metering a sub-cent lookup costs next to nothing. When every HTTP request triggers EIP-3009 verification and on-chain settlement on Base, any endpoint priced below the facilitator's per-transaction fee is uneconomic to sell individually.

Second, variable-cost LLM endpoints break upfront 402 quotes. Brave also excluded its Answers endpoint (/res/v1/chat/completions), which bills $4 per 1,000 requests plus $5 per million input and output tokens on standard plans. An HTTP 402 challenge must quote a fixed amount upfront in the Payment-Required header before execution starts. Fixed-price search routes ($0.005 flat across web search, LLM Context, news, video, image, and three local-search paths) satisfy that contract; token-metered completions whose final bill depends on output length do not.

Third, latency, gas subsidies, and spend governance shift onto the client. Every cold x402 call pays two HTTP round-trips: an unpaid probe returning an empty-body 402 with a challenge valid for 300 seconds, followed by a signed retry. On a search API citing a 669-millisecond average latency in AIMultiple tests on Brave's portal, uncached double round-trips slow down multi-step agent loops unless your client manages challenge timing carefully. Brave currently covers Base gas fees "for a limited time" via its own facilitator, keeping marginal cost at $0.005 today. Once gas coverage ends, effective per-call cost can drift above standard prepaid rates, where accounts also get $5 in monthly free credits and 50 queries per second on Search.

Spend failure modes also change. Brave's verify-before-settle sequence ensures failed upstream 502 searches are never settled, so retrying a broken server call costs nothing extra. What it cannot catch is an agent-logic loop that sends 500 valid HTTP 200 queries while stuck on a task, settling $2.50 on-chain with no vendor billing cap or card dispute path behind it. Your pipeline has to enforce budget discipline inside the wallet layer itself through strict funding limits and flags like purl --max-amount 5000.

For ephemeral CLI tools and fallback routes where human account provisioning stalls deployment, x402 per-request payments remove real setup overhead. For steady production workloads, sub-cent autocomplete calls, and variable-token inference endpoints, prepaid API keys remain the cheaper settlement and governance layer until x402 clients and facilitators support batched session netting.