AI Agents Pricing: Examples & Companies

What is it

AI Agents Pricing is the pricing for AI agent platforms — products that perform multi-step autonomous tasks on the user’s behalf.

Agents break the assumptions older pricing rested on. A chatbot answers one prompt; an agent plans, calls tools, loops, and retries — so a single run can consume 10–100x the tokens of a chat turn and make external tool calls that cost money of their own. Cost per task is therefore high and variable, and the value is the completed job, not the text generated. That mismatch is why the billable unit for agents is still unsettled in 2026.

147 of the in-corpus companies serve the agents use case, and three competing units are in play. Per token is today’s default — Anthropic, OpenAI, and Google price the agent exactly like any other model workload. Per action bills each discrete tool call separately, emerging at agent-native search vendors like Exa and You.com. Per outcome charges only on a verified completion — rare, and led by Intercom Fin at $0.99 per resolution and Sierra’s sales-only per-resolution model.

Most application-layer agents don’t pick one purely. They wrap variable agent work inside a fixed seat or credit allowance — Cursor at $20/mo Pro, Cognition’s Devin at $20 Pro / $200 Max, Rox at $50/mo Core — so the base bill stays predictable while heavy runs draw down (and eventually overflow) the pool.

How it works

Four structures cover almost every agent product in the corpus. The first three are the pure billable units; the fourth is the packaging wrapper most app-layer agents actually ship.

Worked example — the same coding task, three ways. Say an agent completes one feature-implementation task that reads and writes 200,000 tokens on an Opus-class model ($5 input / $25 output per million) and makes 3 web searches.

• Per token: at a rough 150k input / 50k output split, that’s 0.15 × $5 + 0.05 × $25 = $0.75 + $1.25 = $2.00 in model spend, plus 3 × $0.007 = $0.02 for searches at Exa’s $7/1k rate — about $2.02 for the task, and the buyer carries all the variance.
• Per action: the vendor might expose the run as a fixed number of billable actions (a research call, a code-execution call) so the buyer sees legible line items instead of raw token counts — but the underlying token variance still exists, it’s just re-packaged.
• Per outcome: Intercom Fin would charge a flat $0.99 regardless of whether the run burned 40k tokens or 400k — the vendor absorbs the variance, which is exactly why outcome pricing is hard to ship profitably.

Why cost is hard to predict: one agent run might issue 5 tool calls and 40K tokens, while another issues 50 and 400K. Token pricing makes the buyer bear that variance; per-action makes it legible but still variable; per-outcome shifts it to the vendor. The three units are ranked by who bears the cost risk, not by technical difficulty.

This page sits at the intersection of two emerging trends we track: per-action / agent-tool pricing and outcome-based pricing.

Companies using this

147 companies in the current corpus serve the agents use case — the single largest use-case cohort on the site. They span frontier model APIs (Anthropic, OpenAI, Google, DeepSeek), inference clouds (Groq, Together AI, Fireworks AI, Baseten), agent-native search (Exa, You.com, Perplexity AI), coding agents (Cursor, Claude Code, Cognition), support agents (Intercom Fin, Sierra, Decagon, Ada), workflow automation (Zapier, n8n, Lindy), and even embodied agents (1X Technologies). The table below lists each company’s structure, billing units, free tier, and last-verified date.

Patterns observed

Per-token still dominates, but it’s the floor, not the future. Most agent platforms bill tokens because that’s what the underlying model costs — Groq prices Llama 3.3 70B at $0.59/$0.79 per million tokens, and every inference cloud (Together AI, Fireworks AI, Baseten, Replicate) meters the same way. It’s the honest cost-plus floor, but it pushes all consumption risk onto the buyer, which is why nearly no application-layer agent bills raw tokens directly.

Tool calls are becoming their own line item. Beyond the flat search rate, the per-action layer is already fanning into sub-units — Exa adds Deep Search at $12–15/1k and Contents at $1/1k pages, You.com adds Research from $12/1k, and Groq has begun pricing built-in agent tools (web search, code execution) per session on top of tokens. This is the agent-tool pricing trend in motion — the industry starting to name and bill the actions an agent takes, not just the tokens it moves.

Outcome pricing is the talked-about endpoint, and it now has more than one live case. The unit is spreading past its origin: alongside Intercom Fin, Ada (~$1–$3.50/resolution reported), Decagon, Forethought, Maven AGI, and Lorikeet all bill support resolutions as the unit. Outcome pricing clusters in customer support because that’s where the result is measurable and attributable — see outcome-based pricing.

App-layer agents wrap usage in a seat or credit pool. The seat holds revenue steady while the meter recovers variable cost: Cursor runs usage pools up to a $200 Ultra tier, Cognition bills enterprise Devin in ACUs above its self-serve tiers, and Rox meters its Core plan in Agent Actions. This is hybrid pricing applied to agents — the most common landing spot precisely because it splits the variance between buyer and vendor.

Counterexamples & variants

The raw model API prices the agent like nothing special. The counterexample to “agents need new units” is the frontier APIs selling the exact same per-token rate whether you’re building a chatbot or a 50-tool autonomous agent. There’s no agent-specific unit at all, and that’s deliberate: the API sells capability, and the agent logic — along with its cost risk — lives entirely in the buyer’s application. For an infrastructure vendor, refusing to invent an agent unit is the correct move, because it never controls how chatty the buyer’s agent is.

The embodied-agent variant swaps tokens for robot-hours and hardware. The most extreme divergence is physical agents. 1X Technologies sells its NEO humanoid at a published $20,000 outright or $499/mo (six-month minimum), while Agility Robotics and Apptronik meter robot-hours under Robots-as-a-Service contracts. The agents-pricing frame still applies — you’re paying for autonomous tasks completed — but the unit is a physical hour of labor, not a token, which shows how far the “pay for the job” logic can stretch.

Outcome pricing breaks where the outcome isn’t clean. Per-outcome only works where the result is measurable and attributable, which is why it consolidated in support first. It’s much harder for coding or research agents: what’s the billable “outcome” of a Devin run that writes code a human still has to review? That’s exactly why Cognition bills Devin in subscription tiers and ACUs rather than per shipped feature. The variant to watch is whether coding and research agents ever find a defensible outcome unit — until they do, they stay on hybrid credit pools, and the per-outcome column stays dominated by customer support.

What this means for buyers vs vendors

For buyers

Budget per completed task, not per message. Include tool-call outputs in your token estimate and any per-action fees on top — a chatty agent can cost several multiples of a single chat turn, and the worked example above is a floor, not a ceiling. When a vendor offers outcome pricing, scrutinise two things: the outcome definition (Intercom Fin only bills an end-to-end resolution or a configured handoff, not every conversation) and the per-unit price against your cost to do the job another way. For the fundamentals of matching a price to consumption, start with our introduction to usage-based pricing and the breakdown of usage-based pricing models. If you’re evaluating token spend directly, our pricing calculator lets you model model-by-model cost per task before you commit.

For vendors

Pick the unit that matches who should bear consumption risk. Per-token is simplest but pushes all risk to the buyer — fine for an API like OpenAI, dangerous for an app-layer agent where a runaway loop torches your margin. Per-action (Exa, You.com) makes spend legible and is a good intermediate step. Per-outcome (Sierra, Ada) is the strongest value alignment but demands a measurable, defensible result and shifts variance onto you — only ship it where the outcome is clean, which today means support resolutions. Most agent products land on a hybrid credit pool: a fixed fee plus metered usage that splits the variance while the market figures out outcomes. Whichever you choose, get your value metric right first — see choosing the right usage metric.

Explore this theme in the knowledge graph →