AI Startups Shift to Outcome Pricing as Agent Revenue Surges

The short version

As of October 2026, money and revenue in AI are concentrated at the top. The frontier labs are on track for annual revenue in the tens of billions. Below them, a group of application companies is growing fast with pricing that charges for completed work, not for software access. Funding in early Q4 suggests where the next startup opportunities sit: in the tooling that lets companies trust, secure and pay for autonomous agents. The general-purpose agent itself looks less promising for new entrants.

The reports agree on the direction. They don't always agree on the figures, and the gaps matter for anyone using these numbers to plan a company.

The top of the market keeps growing

OpenAI is the largest single number. Sacra estimates the company reached about $70 billion in annualized revenue in September 2026, up from $20 billion at the end of 2025.4 Axios, as relayed by PYMNTS, reported a run rate of almost $70 billion and said business-to-business revenue had more than doubled since the start of the third quarter.5 Another account frames $70 billion as OpenAI's forecast for the end of 2026, up from an earlier $50 billion estimate. It says the figure was shared with investors as the company seeks more than $30 billion at a $1.4 trillion valuation.10

These accounts conflict. One describes $70 billion as a run rate already reached, the other as a year-end target. Whichever is right, OpenAI's revenue has grown roughly threefold or more within a single year.

Anthropic is growing at a similar pace. The company said its run-rate revenue passed $47 billion in May, and Bloomberg put it above $65 billion by the end of July.1 Its valuation went from $183 billion in September 2025 to $965 billion in May 2026.1 One report notes that OpenAI and Anthropic calculate annualized revenue differently, so comparing them directly is unreliable.10

The capital behind these companies is just as concentrated. Startups raised $510 billion in the first half of 2026, and OpenAI and Anthropic took $217 billion of it, about 43%.13 The bill for that spending is also large. OpenAI reported a 33% gross margin, held down by inference costs. The Financial Times reported projections of $278 billion in cumulative negative free cash flow from 2026 through 2030.4

This is the key point for founders. The labs are growing quickly, but they are not yet profitable businesses, and every company building on top of them takes on some of that cost.

How application companies make money

The more useful lessons for startups come from the layer below the labs. Cursor's annualized revenue reportedly rose from about $100 million to about $4 billion in 16 months before SpaceX agreed to buy it in an all-stock deal worth about $60 billion.3 ElevenLabs grew from $100 million in annualized revenue in January 2025 to $2 billion by February 2026.3 Harvey reported about $190 million in annual recurring revenue at an $11 billion valuation in March.3 By August it had reportedly passed $350 million and was raising at $15.5 billion.6

Mistral is the main example outside the US. The Financial Times reported its run rate above $400 million, about 20 times higher than a year earlier, and CEO Arthur Mensch has set a goal of passing $1 billion by the end of 2026.7

These companies share a pattern: each one owns a specific workflow. The workflows are writing code, generating voice, doing legal work and serving European enterprise customers who prefer non-US models. CRV describes vertical AI's advantage as the workflow depth and data that general-purpose products cannot easily reproduce.6 My conclusion from the revenue figures is that the strongest application businesses are not thin wrappers around a model. They become the place where a particular kind of work gets done.

Perplexity is a useful example because estimates of its revenue vary. The FT reported its annual recurring revenue above $450 million in March 2026 after it moved toward agents and usage-based pricing.2 A later report put annualized revenue at $750 million by August, linking the gain to its Comet agent products and usage pricing.3 If both figures are accurate, a consumer search product grew much faster once it started charging for agent work.

Pricing is now a product decision

The largest change in strategy is pricing. Sierra, co-founded by Bret Taylor, mostly charges per resolved customer interaction. It does not charge per seat. Third parties estimate about $1.50 per resolution, and Sierra does not publish its rates.12 The company reportedly took seven quarters to reach $100 million in annual recurring revenue and only two more to reach $200 million. In May it raised a $950 million Series E at a $15.8 billion valuation.312

Other companies are adopting similar models. Intercom's Fin agent charges $0.99 per resolution. HubSpot moved its Breeze agents to outcome pricing in April at $0.50 per resolved conversation and $1 per recommended lead.15 Vida announced on October 8 that customers can pay per qualified lead or transfer. A company spokesperson gave an example: an agent that handles 500 conversations and produces 20 qualified leads bills for the 20 leads.15 Salesforce's Agentforce takes a middle position. It combines Flex Credits, priced at $500 per 100,000 credits, with per-conversation and per-user options. Analysts describe that as a hybrid model, not a pure outcome model.11

The commentary agrees on when outcome pricing works and disagrees on how widely it applies. Several analyses say it works best when the result is unambiguous. Support tickets are either resolved or not, which is harder to say about a coding task.1216 Cognition's Devin bills per Agent Compute Unit whether or not a task finishes. One critic calls that partial billing presented as a clean metric.16 One guide for bootstrapped founders recommends usage or hybrid pricing, arguing that Sierra's approach depends on enterprise sales teams and long contracts.18

The margin calculations explain the caution. CRV works through an example in which an agent action costs about 54 cents to deliver. At $0.99 per outcome, a successful first attempt leaves about 45% gross margin, and needing two attempts per success means losing money.13 CRV also notes that token prices fell about 98% while enterprise AI bills tripled, because agentic workflows can use up to 30 times as many tokens per task as a chat response.13 One industry guide puts typical agent gross margins at 50–60%, compared with 80–90% for traditional SaaS.14

My view is that outcome pricing has moved beyond marketing and become a go-to-market tool, but only for companies that can measure success and absorb failed attempts. For everyone else, the practical default is a platform fee plus one metered unit with a spending cap.1317

Where early Q4 funding is going

Funding in October points to where new companies might find openings. On October 1, Armadin, founded by Mandiant co-founder Kevin Mandia, raised a $255.5 million Series B for offensive security that uses swarms of agents. One tracker valued it at more than $2.5 billion.22 Another feed listed the valuation as "over $25 billion," which appears to be a mistake.23 Reco raised $55 million days earlier, in what TechCrunch called a crowded agent-security market.8 Rein Security added a $25 million Series A focused on agent security.28 Arena raised $200 million at a $3.1 billion valuation to evaluate AI systems. One analysis said enterprises want independent evidence before giving agents access to sensitive systems.30

Taken together, these rounds show demand for agent security and evaluation as standalone categories. That demand exists because companies are deploying agents faster than they can audit them.

Infrastructure rounds at the seed stage follow the same pattern. Monid raised $7.7 million to act as the transaction layer between agents and external tools.26 Zeroset raised a $5.2 million pre-seed for agent memory.29 Photon raised $4.5 million to run agents inside iMessage and WhatsApp.22 Vertical deals keep coming as well, including mortgage agents (Vesta, $30 million) and restaurant workflow automation (Presto, $10 million).2329

Large bets on consumer and general agents are still happening, but mostly for companies that already have an advantage. Instinct raised a $1 billion Series C at a $10 billion valuation weeks after launching an invite-only consumer agent in August.2225 Manus raised more than $500 million to rebuild as an independent company after its acquisition by Meta fell through.28 DeepSeek is reportedly close to raising at least $12 billion ahead of a planned 2027 IPO.27 These are not deals a new founder can easily replicate.

What this means for founders

My reading of the current market comes down to three points.

1. Own one workflow and its data. The fastest-growing application companies each own one task completely. They don't spread a general assistant across many tasks.36
2. Price based on what you can measure. Charge per outcome where success can be verified. Otherwise use a hybrid model, and set prices that cover failed attempts.1315
3. Sell trust. Security, evaluation, memory and transaction infrastructure for agents are attracting funding because buyers need them before they can deploy agents widely.830

The risk is still the economics. Sequoia's David Cahn now estimates that AI needs $1.5 trillion in end-customer revenue to pay back a single year of hyperscaler capital spending.1 The startups most likely to survive any downturn are those whose revenue is tied directly to work they have completed and verified.