We killed the subscription: why agents should pay per call

An agent uses a tool for one task and moves on, so it should pay for that task and nothing else. A person signs up for a tool, learns it and opens it every day; a monthly plan fits that habit. An agent needs a scraper for one job, a voice model for the next, a price feed for the one after, and may not touch any of them again for a week. Put that agent on entry-tier plans and you pay for six tools all month to use each for a few minutes. Below is what that costs with real list prices, where the break-even sits, and the cases where a plan is still the right answer.

The subscription stack, with prices

Take a modest research agent. It searches the web, scrapes a few pages, reads out a summary, runs the occasional scraper actor and checks a crypto price. Five tools, each on its cheapest paid tier, as of October 2026 and billed monthly:

• Tavily Project: $30 for 4,000 credits (Tavily pricing).
• Firecrawl Hobby: $19 for 5,000 credits (Firecrawl pricing).
• ElevenLabs Starter: $6 for 30,000 credits, about 30 minutes of speech (ElevenLabs pricing).
• Apify Starter: $19, which is $19 of prepaid usage at $0.20 per compute unit (Apify pricing).
• CoinGecko Basic: $35 for 100,000 credits (CoinGecko pricing).

Total: $30 + $19 + $6 + $19 + $35 = $109 a month, before the agent has done anything. Each plan also means an account, a card on file, a key to store and rotate, a quota that resets on the provider's calendar, and one more invoice. Apify says it plainly: unused prepaid usage expires at the end of the billing cycle.

Source: Provider pricing pages · Each figure is the provider's own listed monthly price; see sources.

View data

Item	Value
CoinGecko Basic	$35
Tavily Project	$30
Firecrawl Hobby	$19
Apify Starter	$19
ElevenLabs Starter	$6

The same month, priced per call

Now give that agent a plausible month's work and price it at each provider's own unit price. The usage is an assumption, stated so you can swap in your own; the unit prices are not.

• 300 Tavily basic searches. A basic search is 1 credit (Tavily credits) and pay-as-you-go is $0.008 a credit: 300 × $0.008 = $2.40.
• 200 Firecrawl scrapes. A scrape is 1 credit; Firecrawl has no pay-as-you-go, so the closest unit price is its extra-credit rate on Hobby, 1,000 credits per $5, or $0.005 each: 200 × $0.005 = $1.00.
• 5 minutes of ElevenLabs speech at the Starter extra-minute rate of about $0.20: 5 × $0.20 = $1.00.
• 8 Apify compute units at $0.20: 8 × $0.20 = $1.60.
• 1,500 CoinGecko price calls over x402 at $0.005 USDC each (CoinGecko x402): 1,500 × $0.005 = $7.50.

Total: $2.40 + $1.00 + $1.00 + $1.60 + $7.50 = $13.50. The plans cost $109 for the same work, eight times more, and the agent used 7.5% of its Tavily credits, 4% of Firecrawl's, 17% of ElevenLabs', 8% of Apify's and 1.5% of CoinGecko's.

Source: Provider pricing pages · Computed: usage × unit price from each provider's pricing page or credit docs; usage is a stated assumption.

View data

Item	Value
CoinGecko, 1,500 calls	$7.50
Tavily, 300 searches	$2.40
Apify, 8 CUs	$1.60
Firecrawl, 200 scrapes	$1
ElevenLabs, 5 min	$1

Where the plan wins: the break-evens

Per call is not always cheaper. A flat plan beats per-call pricing once usage passes the point where the plan's price divided by the unit price equals the calls you make. From the list prices:

• CoinGecko: $35 ÷ $0.005 = 7,000 calls a month. Above that, Basic is cheaper than x402.
• Tavily: $30 ÷ $0.008 = 3,750 credits. Above that, Project is cheaper than pay-as-you-go, and it includes 4,000.
• ElevenLabs: $6 ÷ $0.20 = 30 minutes, which is exactly what Starter includes.

So a provider's entry tier is priced to win at roughly its own included quota. The question is not "which is cheaper" but "will this agent use the quota". For a tool the agent hits every hour, yes, buy the plan. For the five tools it touches a few hundred times a month, no. Most agents have one or two tools in the first group and a long tail in the second, and the long tail is what per-call pricing is for.

Some providers have already made the move. Exa has no subscription at all: $4 per 1,000 searches, $10 of free credit a month, "no minimum spend" (Exa pricing). Deepgram starts at $200 of free credit and then charges $0.0077 a minute for streaming transcription with no monthly plan required (Deepgram pricing).

Why cards could not do this, and what changed

The reason APIs sold subscriptions was never that usage pricing is hard to meter. It was that collecting $0.005 is impossible on a card. Stripe's US rate is 2.9% + 30¢ per successful charge (Stripe pricing, as of October 2026): on a $0.005 charge the 30¢ alone is 60 times the amount, and on a $1 charge the fee is 32.9%. A plan at $19 a month collects the fee once and amortises it.

Two things changed. Stripe's machine payments now accept per-call payments from agents with a $0.50 minimum on cards through Shared Payment Tokens and a 0.01 USDC minimum on stablecoins. And x402 revived HTTP 402 Payment Required so a server can price a single request and the agent can pay it in the same exchange, with no account, no key and zero protocol fees (reference implementation). Coinbase's facilitator reports more than 100 million x402 payments across Base and Solana (CDP docs).

Settling every call onchain is still the wrong design for most agents, though: a prepaid balance with a per-call ledger gives the same economics without a wallet signature per request. That is the model AMXP uses: one balance, a reservation before the call, the exact amount settled after, the hold released if the provider fails.

You still set the limits

Paying per call is not an open tab. Three controls do the work the plan's quota used to do, and do it better, because they are yours rather than the provider's:

• A ceiling per call. The agent cannot run anything whose inspect price is above it, which stops a single expensive endpoint (a 250-credit research call, say) from being chosen by accident.
• A cap per month. The balance will not go below it no matter how many loops the agent runs; the five-plan stack above has no equivalent, because each plan's overage is set by the provider.
• An auto top-up threshold. Set it above a day's normal spend so a provider outage that triggers retries does not drain and refill the balance while you sleep.

Every call lands in one ledger with the tool, the parameters, the reserved amount and the settled amount, so the question "what did the agent spend on Tuesday" has a one-line answer instead of five dashboards.

What most people get wrong

1. Comparing the plan's per-credit price to the per-call price. $35 for 100,000 CoinGecko calls is $0.00035 a call, cheaper than x402's $0.005, but only if you make 100,000 calls. Divide the plan by the calls you actually make.
2. Counting only the price. Each plan is also a key to rotate and a quota that resets on its own date. Five plans are five incidents waiting for the day a key leaks or a quota runs out at 2 a.m.
3. Letting the agent spend without a ceiling. Pay per call needs a cap per call, a cap per month and a top-up threshold, or one bad loop empties the balance. Set them before the agent runs.
4. Putting the hourly tool on per-call pricing. If the agent calls one API 50,000 times a month, buy that plan and keep per-call for the rest.
5. Trusting a hard-coded price. CoinGecko says its x402 pricing "may change without notice". Read the price from the 402 or from inspect on every call.

Common questions

Is pay-per-call cheaper than a subscription?

Below the plan's break-even, yes. For CoinGecko that is 7,000 calls a month; for Tavily, 3,750 credits. Above it, the plan is cheaper per call. Most agents are below it on most tools and above it on one or two.

Why not just give the agent my API keys?

For a tool you already pay for and use heavily, do. AMXP is for the long tail: tools the agent needs once a week, ones you would never sign up for, and ones it has not discovered yet. A good agent checks your own tools first and comes to AMXP for the gaps.

How does AMXP charge?

AMXP passes on the provider's price plus a small, visible fee per call, and the agent sees the total in inspect before it runs anything. See pricing for balances, reservations and top-ups.

AMXP is in private preview. Balances, holds and spend controls work as described once accounts open; get early access.