The Agent Economy: Paid Research and Data for AI Agents, 2026 to 2027

Cheaper tokens, open payment rails and the small volume that still has to prove itself

The price businesses pay per token fell hard in 2026, through both price cuts and a move to cheaper models that public series cannot tell apart, and the card networks, Stripe, Coinbase and Cloudflare have each put rails in place for software agents to pay per request. The money that public trackers can see moving is still tiny and partly artificial, and it is not yet clear whether data owners or the agent platforms that choose which sources to pay will keep the margin.

ISSUED 12 OCTOBER 2026 · ECONOMIC & FINANCIAL MARKETS · AEGEAN INTELLIGENCE GROUP · AUDITED AGAINST AEGEAN-OS

THE DOCUMENT

NumberAIG-BLK-26-008

Product IDPMR-2026-1012-GLB-008

ProductAssessment

Issued2026-10-12

Pages18

DistributionPublic release

TimelinessCURRENT

DOWNLOAD

FULL PUBLICATION (PDF)

18 PAGES · 0.5 MB · FREE

HOW TO CITE

Aegean Intelligence Group, “The Agent Economy: Paid Research and Data for AI Agents, 2026 to 2027,” AIG-BLK-26-008, The Aegean Register, 12 October 2026.

The Agent Economy: Paid Research and Data for AI Agents, 2026 to 2027 cover

Abstract

Central finding

Two costs that kept research a human, subscription business are falling at once: the price of the tokens that generate an answer and the cost of charging for a single request. The token price fall is real, driven by price cuts and by buyers moving to cheaper versions, mini models and open models in proportions no public series can split, but agentic research consumes many more tokens per task, so cheap tokens do not yet mean cheap finished research. Where both costs are low, trusted inputs such as licensed data, primary documents, proprietary series and verifiable research become relatively scarcer than analysis, but the agent platforms, content networks and payment gateways that decide which sources an agent calls are positioned to set terms too. Most money agents spend on data today still moves through API keys, enterprise contracts and platform billing, and the agentic checkout the card networks are building runs on tokenised cards, so the open-standard volume this paper measures is a narrow window on the market rather than its size. The payment rails are spreading faster than demand, so the near-term market is shaped by sellers of data and research, by the platforms that buy on agents' behalf, by the rules that make dollar stablecoins usable by regulated firms, and by whether real buyers replace speculative and test traffic.

Publication at a glance

Source registry23 entries

Forward claims3

Pages18

Issued 2026-10-12Public releaseCURRENTAIG-BLK-26-008

Cover

On the record

3 calls from this report, among 49 in Economic & Financial Markets.

Each line is one call in the domain, from the day it was written to the day it resolves. This report's calls are lit; point at any line to read it.

FORWARD CLAIMS · DATED TO PUBLICATION · SCORED ON THE SCORECARD →

ID	Statement	p	Made on	Resolves by	Status
CLM-2026-409	The Office of the Comptroller of the Currency publishes a final rule implementing the GENIUS Act for OCC-supervised payment stablecoin issuers in the Federal Register on or before 10 January 2027.	0.65	2026-10-12	2027-01-10	Locked
CLM-2026-410	The Federal Deposit Insurance Corporation publishes a final rule implementing the GENIUS Act requirements and standards for FDIC-supervised permitted payment stablecoin issuers in the Federal Register on or before 10 January 2027.	0.40	2026-10-12	2027-01-10	Locked
CLM-2026-411	The total stablecoin market capitalisation on the DefiLlama stablecoins dashboard is at or above 300 billion USD for 31 December 2026.	0.65	2026-10-12	2027-01-05	Locked

A claim is entered in the brief's own words and carries its publication date and the day it entered the ledger: the register's briefs on 8 September 2026, when the standard was first applied; the forward benchmark on the day it was locked. It resolves only against a dated public source. The scoring rule is on the Methodology page.

THE CENTRAL JUDGMENT

KEY JUDGMENTS

Bottom line

KJ-01The effective price per token paid by business users fell sharply in 2026, and both price cuts and trading down to cheaper models drove the fall. These usage-weighted series cannot split the two drivers, since moves within a tier to cheaper versions and mini models and to cheap open models count as mix just as a shift away from frontier models does, so we make no call on which driver was larger. Two independent series, the spending of businesses on Ramp's platform and the Silicon Data index, show the same steep fall from different data.^3,4HIGH

KJ-02The effective price per token is more likely to stay below Ramp's March peak of 1.15 USD per million tokens and the Silicon Data May high of about 2 USD per million tokens than to regain them through mid-2027. Both series are short, mix-weighted and volatile, rising sharply earlier in 2026 before falling, so a new frontier release or a swing of agentic work back to frontier and reasoning models could lift the effective price back toward its peak within months, but the summer run of price cuts and cheap open models makes a full return the less likely path.^3,4,22MODERATE

KJ-03Pay-per-request rails for agents are now available and broadly backed by the main card networks, Stripe, Coinbase and Cloudflare, though not yet standard in deployment. For most sellers the choice of rail is likely to matter less than the ability to price and prove what they sell. The rail can still matter in two ways: the largest agent platforms, not sellers, choose which rails they call, so a seller on a rail they shun would not be called. It also sets a floor on price, because card rails carry fixed per-transaction fees that make a sub-dollar call uneconomic unless it is batched, while stablecoin rails settle small sums cheaply. The same payment firms sit as members or partners across x402, the Machine Payments Protocol and Mastercard's offer built for small transfers, and gateways such as Cloudflare's put one integration in front of several rails, which lowers the cost of a wrong choice.^1,7,8,15MODERATE

KJ-04Real commercial volume from agents buying data and research stays below one hundredth of one percent of card network payment flows through mid-2027, measured as a monthly ratio against Visa's payments alone, about 1.2 trillion USD a month on its fiscal 2025 total, which puts the bar near 120 million USD a month and would put it higher if other networks were counted. Off-chain volume on API keys, enterprise contracts, platform billing and tokenised card checkout is judged only from what firms disclose publicly, and the judgment fails if a disclosed figure for agent purchases of data and research in any month passes the bar. Total measured x402 volume was about 0.002 percent of Visa's monthly payments on a single 30-day homepage reading in July 2026, and a third-party tracker put cumulative settled x402 volume near 42 million USD by October 2026, a total hard to square with that July reading as a steady monthly rate, so on-chain purchases of data and research sit well below the bar.^1,20,23MODERATE

KJ-05Owners of licensed, proprietary and verifiable data are better placed than sellers of generic answers to keep a price from agents, but the agent platforms and gateways that choose which sources to pay may take much of the margin. The one disclosed licensing price is a single Wall Street Journal figure carried second hand in a Fortune newsletter, and the buyer-set terms of the newest per-use program rest on an UNVERIFIED Cloudflare blog read in search extracts, so the evidence fits a data-owner story and an aggregator story about equally.^13,14LOW

KJ-06The OCC is more likely than not to publish its final GENIUS Act issuer rule by early January 2027 and the FDIC slightly less likely than not, but regulated banks and large data vendors are unlikely to take dollar stablecoins from agents in amounts that matter to their revenue by early 2027, whether or not the rules are final; switching on processor-settled stablecoin acceptance may well spread in that time, while material receipts will not. The Comptroller's November target, carried by Decrypt and by one Forkast report republished by Yahoo Finance, and the hard 18 January 2027 effective date push the OCC toward the window, while measured agent payment volume is far too small to be material to a large bank's or data vendor's revenue.^1,2,16,17MODERATE

SOURCES AND LIMITS

Open-source intelligence current through 12 October 2026. The publication carries a 23-entry source registry, explicit confidence tags, limitations, and deterministic resolution rules for its forward claims.

SOURCE REGISTRY: 23 ENTRIES18 PAGESTIMELINESS: CURRENT

The source tiers, estimative language, forward-claim rule, scoring, and corrections record are public on the Methodology page.