{
  "data": {
    "author": "Quynh Tran-Thanh",
    "faq": null,
    "kicker": "Blog · From the data · 10 November 2026",
    "lede": "A fifth of the agent stack still sells seats, and none of the A-graded listings do. I went through every listing, a year of shutdowns and the public numbers to see which business models are growing, which are being bought or switched off, and what survives next to a lab.",
    "readingMinutes": 14,
    "words": 3126
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/who-gets-paid-in-the-agent-stack",
    "json": "https://www.anchorterminal.com/blog/who-gets-paid-in-the-agent-stack.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/who-gets-paid-in-the-agent-stack.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/who-gets-paid-in-the-agent-stack.min.md"
  },
  "markdown": "By Quynh Tran-Thanh, 10 November 2026. 3,126 words, about 14 min to read.\n\nEveryone I talk to has a version of the same two worries. Is SaaS over now that agents do the clicking, and will the labs eat whatever is left? I wanted an answer with numbers in it, so I went through the 1,644 listings on Anchor Terminal, the investor records we keep for each one, every shutdown notice in the catalogue, and a pile of public filings and surveys. This is what I found, with the caveats where they belong.\n\n## How the stack charges today\n\nEach listing records how the vendor charges. Across the catalogue on 10 November 2026 the split is freemium 36 per cent, free 23 per cent, seat or plan pricing 20 per cent, metered usage 19 per cent, and 2 per cent bring your own plan, where the client is free and you pay the underlying vendor [[1]](#ref-1).\n\n- 593: freemium, a free tier and then plans\n- 372: free, open source or no charge\n- 333: seats or plans, no free tier\n- 308: metered usage\n- 38: bring your own plan\n\nHow 1,644 listings charge, 10 November 2026. Source [1].\n\nThe shape depends on what the thing costs to make. Anything whose cost of goods is compute or a third party's per-unit fee meters. Inference is 62 per cent usage priced, GPU compute 76 per cent, voice calling 87 per cent, crypto exchanges 90 per cent [[1]](#ref-1). Anything bought by a human team sells seats. CRM is 70 per cent seat priced, support 64 per cent, company knowledge 65 per cent, recruiting 80 per cent, and tax, payroll and HR are 100 per cent [[1]](#ref-1). Frameworks, SDKs, harnesses and open weights are almost all free, which makes sense once you see them as distribution for a metered thing one layer down. The MCP layer barely charges at all. Of 163 MCP servers, 7 per cent are usage priced and 14 per cent sell plans [[1]](#ref-1).\n\nUsage against seat pricing in the categories at each end, per cent of listings\n\n| Category | Metered usage | Seats or plans |\n| --- | --- | --- |\n| Crypto exchanges | 90 | 0 |\n| Voice calling | 87 | 7 |\n| GPU compute | 76 | 5 |\n| Inference | 62 | 5 |\n| CRM | 0 | 70 |\n| Support | 0 | 64 |\n| Company knowledge | 0 | 65 |\n| Recruiting | 0 | 80 |\n| Payroll, tax and HR | 0 | 100 |\n\nEach series has a zero-based scale.\n\nThe middle of the catalogue is freemium. Connectors, observability, memory, scrapers, search and agent auth are 50 to 86 per cent freemium [[1]](#ref-1). My reading is that these vendors have not settled what their unit is yet, so they give the product away and charge for the plan around it.\n\n## The top tier does not sell seats\n\nOur grade measures quality and transparency against the benchmark, not traction. But the pricing mix moves with it in a way I did not expect to be this clean.\n\nHow each tier charges, per cent of listings in the tier\n\n| Tier | Metered usage | Freemium | Free | Seats or plans |\n| --- | --- | --- | --- | --- |\n| Grade A and AA, 28 listings | 44 | 24 | 32 | 0 |\n| Top 50 by score | 44 | 28 | 26 | 2 |\n| Top 250 | 28 | 42 | 22 | 8 |\n| Grade C, 420 listings | 14 | 34 | 30 | 20 |\n| Grade F, 141 listings | 9 | 34 | 15 | 40 |\n\nEach series has a zero-based scale.\n\nSeat pricing goes from none of the A grades to 40 per cent of the F grades, and metering runs the other way [[1]](#ref-1). The top 50 is 24 metered APIs (Vertex AI, the OpenAI API, Bedrock, Stripe, Twilio, S3), 13 free frameworks and protocols (the AI SDK, the OpenAI Agents SDK, Pydantic AI, x402) and 12 freemium infrastructure products (Infisical, Resend, Apify, Tavily). One seat product [[1]](#ref-1).\n\nPart of this is the benchmark itself. Payments and transparency reward a published unit price, which metered vendors have and seat vendors tend to hide behind a sales call. So read the table as \"the stack an agent can use well is metered or free\", not as \"usage pricing makes you good\". I think both are true, but the data only shows the first.\n\n## Where the money went\n\nWe record named investors and dated rounds for each listing. 235 of the 1,644 had a round dated November 2025 or later [[2]](#ref-2). By pricing model, 25 per cent of usage-priced listings raised in the last twelve months, against 14 per cent of freemium, 12 per cent of free and 9 per cent of seat-priced ones [[2]](#ref-2). The seat-heavy categories, CRM, support, CMS, commerce, scheduling, recruiting, payroll, had no rounds at all in the window [[2]](#ref-2).\n\nListings with a round dated November 2025 or later, by pricing model, per cent\n\n| Pricing model | Per cent raised |\n| --- | --- |\n| Metered usage | 25 |\n| Freemium | 14 |\n| Free | 12 |\n| Seats or plans | 9 |\n\nEach series has a zero-based scale.\n\nTwo qualifications. Usage-priced categories (inference, GPU, voice, video) are capital hungry, so some of that funding is capex rather than growth. And the fourth most frequent investor across the catalogue is NVIDIA, on 98 listings, behind Y Combinator, a16z and General Catalyst [[2]](#ref-2). The supply chain is funding its own demand.\n\n## Who is growing\n\nIn public markets the split is by meter, not by sector. Snowflake grew product revenue 37 per cent in the quarter to 31 July 2026 [[3]](#ref-3), Cloudflare 39 per cent in the quarter to 30 September [[4]](#ref-4), Datadog 35 per cent in the same quarter [[5]](#ref-5). Salesforce grew 11 per cent [[6]](#ref-6), Workday's subscription revenue 14 per cent [[7]](#ref-7), Zoom 5 per cent [[8]](#ref-8). The exception that keeps me honest is Figma, seat priced with AI credits bundled in, up 43 per cent with an 80 per cent gross margin, and even there the non-GAAP margin fell four points in a year as it absorbs agent inference [[9]](#ref-9).\n\nYear on year revenue growth, per cent, as each company reports it\n\n| Company | Growth |\n| --- | --- |\n| Figma, seats with AI credits | 43 |\n| Cloudflare, metered | 39 |\n| Snowflake, metered, product revenue | 37 |\n| Datadog, metered | 35 |\n| Workday, seats, subscription revenue | 14 |\n| Salesforce, seats | 11 |\n| Zoom, seats | 5 |\n\nEach series has a zero-based scale.\n\nAmong private companies, every one above 100 per cent growth with a dated figure is usage, credits or outcome priced, or sells inference. Modal said in October that it had passed a $1 billion run-rate, about 15 times a year earlier [[10]](#ref-10). Sacra estimates Exa at about $100 million annualised in September against $10 million a year before [[11]](#ref-11), and Turbopuffer at about $100 million on less than $1 million raised [[12]](#ref-12). Lovable said it crossed $500 million in annualised revenue in June and $600 million in September, selling credits [[13]](#ref-13) [[64]](#ref-64). Sacra estimated Supabase at $170 million ARR, up about 221 per cent year on year, around its $500 million raise in June [[14]](#ref-14).\n\n## Is SaaS going to survive?\n\nYes, as an ex-growth model, and in specific places.\n\nSeats hold where the buyer is a human team and the product is a system of record or replaces billable labour. Harvey said on 1 October that it had passed $400 million in ARR with 97 per cent gross retention [[15]](#ref-15), a month after raising $550 million at $15.5 billion [[60]](#ref-60). It is also the sharpest case of the margin problem below. Bloomberg reported that its gross margin went from about 50 per cent in January to about minus 50 per cent by June, as agent token use rose twentyfold on fixed-price seats, and recovered only after it shipped its own model in August [[59]](#ref-59). Harvey has not changed how it charges. Its forward-deployed pods, a product manager, a lawyer or two and engineers inside each large client, are a delivery motion sold with the seats [[62]](#ref-62), and its October positioning as the orchestration and connector layer for legal work is a platform claim, not a price change [[61]](#ref-61). Figma is the public version of the same story. In our catalogue the seat-priced verticals did not shut down this year. They also did not raise.\n\nA year of seats and token costs, as reported\n\n1. **January 2026** (margin positive). Gross margin about 50 per cent on fixed-price seats [[59]](#ref-59)\n2. **Spring** (costs rising). Agent token use rises twentyfold, the seat price does not move [[59]](#ref-59)\n3. **June** (margin negative). Gross margin about minus 50 per cent [[59]](#ref-59)\n4. **August** (margin positive). Harvey ships its own model and the margin recovers [[59]](#ref-59)\n5. **September** (price unchanged). Raises $550 million at a $15.5 billion valuation, still selling seats [[60]](#ref-60)\n6. **1 October** (price unchanged). Says it has passed $400 million in ARR with 97 per cent gross retention [[15]](#ref-15)\n\nThe surveys say the same thing from the other side. In Kyle Poyar's May 2026 survey of 230 companies, hybrid pricing is the most common primary model at 37 per cent, up from 25 per cent a year earlier. Asked which model they prefer, investors put hybrid at 35 per cent, outcome at 26 per cent, usage at 24 per cent and seats at 5 per cent [[16]](#ref-16). ICONIQ's July survey has consumption pricing rising from 35 to 42 per cent of companies and outcome pricing from 18 to 23 per cent, with subscriptions still used by 57 per cent [[17]](#ref-17).\n\nWhich pricing model investors say they prefer, per cent\n\n| Model | Investors preferring it |\n| --- | --- |\n| Hybrid | 35 |\n| Outcome | 26 |\n| Usage | 24 |\n| Seats | 5 |\n\nEach series has a zero-based scale.\n\nBenchmarkit puts median net revenue retention at 108 per cent for usage-based companies and 98 per cent for seat-based, from a 230-company cut whose usage-based subset it does not size [[18]](#ref-18).\n\nSo hybrid, not a clean switch. Incumbents bolt credits onto seats. AI natives add seats and platform fees for predictability. Credits are what both sides converged on after bill shock. Replit moved its agent to effort-based pricing in June 2025 [[19]](#ref-19), Cursor moved teams to seats with two usage pools in June 2026 [[20]](#ref-20), and in Poyar's survey 29 per cent of companies already sell AI credits with another 33 per cent planning to [[16]](#ref-16).\n\nThe reason none of this is comfortable is margin. Bessemer puts AI companies at 50 to 60 per cent gross margin against 80 to 90 per cent for classic SaaS [[21]](#ref-21). The companies that have made usage pricing work either own inference, pass tokens through, or sell the meter itself. Stripe agreed in August to buy OpenRouter, which charges a 5.5 per cent fee on credit purchases and no markup on inference, for a reported $7.5 billion, and the deal had not been announced as closed when I wrote this [[22]](#ref-22) [[65]](#ref-65). Owning the billing and choice layer is now valued above owning a model.\n\n## Are the labs eating the market?\n\nPartly, and in one specific place. I went looking for every company in the stack that shut down, was acquired or sunset a core product between November 2025 and November 2026, and found 47. Twenty-four were bought by an incumbent or another startup, six were labs retiring their own products, five shut or pivoted on their own, and twelve were deals by a frontier lab. Where the target sold a hosted product in the thin layer next to the model, it was switched off within weeks. Anthropic bought Vercept and shut its computer-use agent 28 days later [[23]](#ref-23), bought Stainless and wound down all of its hosted products, SDK generator included [[24]](#ref-24), and acqui-hired Mendral, whose founders wrote that every few months a new model made part of their roadmap unnecessary and a bigger part of it possible [[25]](#ref-25). Meta bought Conductor on 4 November and its cloud closes on 20 November [[27]](#ref-27). OpenAI took longer with Ona, bought in June, with Ona Cloud shut on 31 October [[26]](#ref-26) [[63]](#ref-63). Where the target was open source the lab kept it running. Promptfoo stayed open and supported after OpenAI bought it [[44]](#ref-44).\n\nThe 47, by who did it\n\n| Who | Companies or products |\n| --- | --- |\n| Bought by an incumbent or another startup | 24 |\n| Bought by a frontier lab | 12 |\n| A lab retiring its own product | 6 |\n| Shut or pivoted on its own | 5 |\n\nEach series has a zero-based scale.\n\nHow long a lab kept the hosted product running after buying it\n\n| Who | Time | Outcome |\n| --- | --- | --- |\n| Conductor, to Meta | 16 days | cloud shuts on 20 November |\n| Vercept, to Anthropic | 28 days | Vy switched off |\n| Ona, to OpenAI | 142 days | Ona Cloud shut on 31 October |\n| Promptfoo, to OpenAI | still open | the open-source tool kept and supported |\n\nTime on a log scale.\n\nThe pattern in those twelve is the thinnest layers between the model and the developer. Computer use, sandboxes, SDK and MCP generation, evals, prompt management, coding assistants. Each is now a native feature of a lab product. Our catalogue agrees. Labs are 8 per cent of listings overall but a fifth of agent harnesses and half of guardrails [[1]](#ref-1), and the in-catalogue casualties cluster in voice, harnesses and sandboxes. Google is winding down Jules, read-only from tomorrow and gone on 31 December [[28]](#ref-28), PlayHT's platform closed on 31 December after Meta hired its team [[45]](#ref-45), and Hume's voice APIs and Runloop's support both end on Friday [[29]](#ref-29).\n\nThree things complicate the story.\n\nFirst, most exits were incumbents buying, not labs, and the products kept running. Salesforce paid about $3.6 billion for Fin, the company formerly called Intercom, closing in September [[30]](#ref-30) [[47]](#ref-47), and has agreed a reported $2.7 billion for Cresta [[31]](#ref-31). Dynatrace paid $915 million for Arize [[32]](#ref-32). Nebius paid $275 million plus an earn-out for Tavily [[33]](#ref-33). That is consolidation, not extinction, and it rewards the same metered infrastructure the funding data does.\n\nSecond, the labs killed their own agent products at the same rate. OpenAI shut the Assistants API in August and is shutting Agent Builder on 30 November [[34]](#ref-34). ChatGPT Atlas lasted nine months [[35]](#ref-35). Building on a lab's agent surface carried the same platform risk as competing with it.\n\nThird, the market panic was specific and indiscriminate at once. The selloff that began on 3 February, which a Jefferies trader christened the SaaSpocalypse, followed Anthropic's Cowork plugins for legal, finance, sales and marketing work [[36]](#ref-36) [[46]](#ref-46). The IGV software ETF fell 24 per cent in the first quarter, its worst since 2008 [[37]](#ref-37), and in April ServiceNow had its worst day on record with Workday down 45 per cent for the year [[38]](#ref-38). Snowflake grew 37 per cent through the same months. The market sold seats and meters together. The data says it should not have.\n\n## Who thrives next to a lab\n\nThe question I get asked most is the inverse. If you are not the lab, what can you be? The year gives six answers, and the catalogue backs them.\n\nStart with our data. Counting only the six frontier labs (OpenAI, Anthropic, Google, Meta, xAI and Mistral), they have 83 listings in 37 of our 100 categories. In those 37 categories the best-graded listing belongs to a lab in six, to someone else in fifteen, and the two tie in sixteen [[1]](#ref-1).\n\nIn the 37 categories where a frontier lab has a listing, whose listing is graded best\n\n| Who has the best grade | Categories |\n| --- | --- |\n| Someone who is not a lab | 15 |\n| A tie | 16 |\n| A frontier lab | 6 |\n\nEach series has a zero-based scale.\n\nThe labs lead in inference, open weights and advertising. Everywhere else somebody who is not a lab has the better-documented, better-priced tool in the same category. The 63 categories with no lab listing at all are the ones with a long integration tail or a regulated counterparty. Web scrapers, support, lead data, crypto exchanges, commerce, banking data, identity verification, payroll, tax [[1]](#ref-1).\n\nOutside the catalogue the pattern is the one Snowflake and Datadog ran on AWS, and it comes down to a few traits.\n\n- Neutrality the lab cannot offer. A router that sends traffic between labs is worth more than any one lab's router. Menlo, an OpenRouter investor, wrote that it had the highest uptime of even frontier models because it could contract across providers [[49]](#ref-49). The counter-case is Cursor. Once SpaceX owned it, OpenAI said it would stop supplying models, with the shutoff set for 12 November [[50]](#ref-50).\n- A cost of goods the lab does not want to carry. Expert data is people-heavy and low margin. Mercor reached a $2 billion gross run-rate with about 91 per cent of it from OpenAI and Anthropic [[51]](#ref-51). Serving open-weight and custom models is the same trade in compute, which is where Modal and Fireworks sit.\n- Liability and humans in a regulated vertical. Harvey puts a former lawyer in every deployment [[62]](#ref-62), and Anthropic's healthcare lead said of Abridge that there is no need to reinvent the wheel when a product already serves the use case [[52]](#ref-52). Anthropic shipped Claude for Legal in May and Harvey's valuation went up anyway.\n- Data the model cannot train on. The analysts who called the February bottom named FICO, Moody's, S\u0026P Global, Verisk and Thomson Reuters as the beneficiaries [[53]](#ref-53).\n- A modality the labs treat as second priority. ElevenLabs passed $500 million in ARR in the first four months of the year, by its own account, selling by the character and the minute [[54]](#ref-54).\n- Owning the end user and the deployed artefact. Lovable and Replit sell to people who are not developers and host what they build. The labs sell the editor to developers.\n\nWhat does not survive is the mirror image. A model-neutral tool that loses its neutrality, Cursor this year and Windsurf before it, which Anthropic cut off when OpenAI tried to buy it [[55]](#ref-55). A single-step text-out wrapper on the horizontal road the labs' own distribution runs down, which is a16z's phrase for coding, writing and images [[56]](#ref-56). Developer tooling a lab needs internally, which gets bought and switched off. And a hosted copy of something the lab will open source, as Anthropic did with Claude Code's sandbox runtime in October [[57]](#ref-57).\n\n![The lab keeps the model, inference, open weights, advertising and the developer's editor. The thin layer beside the model, computer use, sandboxes, SDK and MCP generation, evals, prompt management and coding assistants, gets bought and switched off. Six positions survive around it. Neutrality across labs, a cost of goods the lab will not carry, liability and humans in a regulated vertical, data the model cannot train on, a second-priority modality, and owning the end user and what they build.](https://www.anchorterminal.com/assets/img/blog/who-gets-paid-in-the-agent-stack/next-to-a-lab.svg)\n\nWhere you can stand next to a lab. The six positions from this year's exits and the catalogue.\n\nSarah Guo's version is the one I keep coming back to. There is only one A-team at each lab, and if your product is the 27th priority for Google it is not that scary [[58]](#ref-58).\n\n## What stays\n\nFrom the exits and the funding together, the categories that neither the labs nor the incumbents have touched share one of three traits. A heavy integration surface, a cost of goods the lab does not want to carry, or a countable outcome.\n\n- Metered infrastructure one layer below the model. Inference clouds, sandboxes, search and browser APIs, databases that agents provision. Menlo's survey has code-sandbox adoption among AI builders going from 11 to 40 per cent in a year and search from 5 to 30 per cent [[39]](#ref-39).\n- Per-minute voice APIs, despite native lab voice. Sacra estimates Retell at about $80 million annualised on $4.6 million raised [[40]](#ref-40). The telephony integration is the moat.\n- Outcome-priced agents with a countable unit. Fin charges $0.99 per outcome, unchanged since Salesforce completed the purchase in September [[48]](#ref-48), and Bessemer had it on track for $100 million in ARR a year ago [[41]](#ref-41).\n- Vertical seat businesses that replace billable labour, Harvey being the case, as long as they can carry the inference cost. Harvey answered that by training its own model, not by changing the price.\n- The metering, billing and choice layer itself.\n- Data and payment rails the labs will not own. In our catalogue these are old, seat or usage priced, and untouched.\n\n- 11% to 40%: AI builders using a code sandbox, in a year\n- 5% to 30%: AI builders using a search API, in a year\n- $0.99: per outcome, Fin's price before and after Salesforce\n- $500M: ElevenLabs' ARR, by the character and the minute\n- ~$80M: Retell's annualised revenue on $4.6 million raised\n\nWhat the survivors sell by. Sources [39], [40], [48] and [54].\n\nWhat goes is a hosted wrapper around a single model capability, a standalone vector store (Chroma Cloud is being wound down after Chroma agreed to join Replit, writes off on 7 December and reads on 22 December [[42]](#ref-42)), a low-code builder (Flowise announced end of life on 29 July, with its maintainer saying coding agents had taken over the job [[43]](#ref-43)), and anything whose roadmap is the next model release.\n\n## What I'd watch\n\nThe categories where our top quartile already charges differently from the rest are the ones mid-migration, and in every case the top is metered while the rest sell plans. Identity verification, SEO, decision models, secrets, document extraction, human in the loop [[1]](#ref-1). If you sell plans in one of those, the vendor above you in the ranking has already switched.\n\n---\n\nMore in [From the data](https://www.anchorterminal.com/blog/from-the-data/index.md) · Previous from the data [The best harnesses for a personal agent, by our grades](https://www.anchorterminal.com/blog/best-harnesses-for-personal-agents.md)\n\n- Older: [A machine that builds the benchmark](https://www.anchorterminal.com/blog/a-machine-that-builds-the-benchmark.md)\n- Newer: [What Sapolsky's first lecture means for LLMs and decision models](https://www.anchorterminal.com/blog/sapolsky-first-lecture-decision-models.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\nAnchor Terminal grades the tools AI agents run on, with a source for every score. [The top tools](https://www.anchorterminal.com/top/index.md) · [Categories](https://www.anchorterminal.com/categories/index.md) · [Compare two tools](https://www.anchorterminal.com/compare/index.md) · [Prices](https://www.anchorterminal.com/prices/index.md) · [Sunsets](https://www.anchorterminal.com/sunsets/index.md) · [API, MCP server and CLI](https://www.anchorterminal.com/agents/index.md)\n\n## References\n\nSources read on 10 November 2026 unless stated.\n\n[1] Anchor Terminal, the category pages and the listing data behind them, 10 November 2026. Source for the pricing mix across 1,644 listings, by category, kind, grade and score tier, for the lab share of listings, and for the unchanged popularity fields since the 3 October data. https://www.anchorterminal.com/categories/\n\n[2] Anchor Terminal, people and investor records on the listing pages, 10 November 2026. Source for rounds dated November 2025 or later by pricing model and category, and for the most frequent investors. https://www.anchorterminal.com/people/\n\n[3] Snowflake, Form 8-K exhibit, second quarter fiscal 2027 results, August 2026. Source for product revenue growth of 37 per cent. https://www.sec.gov/Archives/edgar/data/0001640147/000164014726000033/fy2027q2earnings.htm\n\n[4] Cloudflare, third quarter 2026 financial results, 5 November 2026. Source for revenue growth of 39 per cent. https://cloudflare.net/news/news-details/2026/Cloudflare-Announces-Third-Quarter-2026-Financial-Results/default.aspx\n\n[5] Datadog, Form 8-K exhibit, third quarter 2026 results, 4 November 2026. Source for revenue growth of 35 per cent. https://www.sec.gov/Archives/edgar/data/0001561550/000162828026069792/ex-991x20260930x8k.htm\n\n[6] Salesforce, second quarter fiscal 2027 results, 26 August 2026. Source for revenue growth of 11 per cent. https://www.salesforce.com/news/press-releases/2026/08/26/fy27-q2-earnings/\n\n[7] Workday, fiscal 2027 second quarter financial results, 27 August 2026. Source for subscription revenue growth of 13.9 per cent. https://newsroom.workday.com/2026-08-27-Workday-Announces-Fiscal-2027-Second-Quarter-Financial-Results\n\n[8] Zoom, second quarter fiscal 2027 results, 25 August 2026. Source for revenue growth of 4.9 per cent. https://www.globenewswire.com/news-release/2026/08/25/3350878/0/en/zoom-communications-reports-financial-results-for-the-second-quarter-of-fiscal-year-2027.html\n\n[9] Figma, Form 8-K exhibit, third quarter 2026 results, 4 November 2026. Source for revenue growth of 43 per cent and GAAP gross margin of 80 per cent. https://www.sec.gov/Archives/edgar/data/0001579878/000162828026070133/q326pressrelease.htm\n\n[10] The Next Web, Modal raises $750M at $15.75bn valuation as run rate tops $1bn, 14 October 2026. Source for Modal's stated run-rate and growth at its Series D. https://thenextweb.com/news/modal-750m-series-d-15-75bn-valuation\n\n[11] Sacra, Exa company profile, September 2026. Source for Sacra's estimate of Exa's annualised revenue and the year-earlier figure. https://sacra.com/c/exa/\n\n[12] Sacra, Turbopuffer company profile, March 2026. Source for Sacra's estimate of Turbopuffer's annualised revenue and capital raised. https://sacra.com/c/turbopuffer/\n\n[13] TechCrunch, Lovable says it has hit $500M in annualized revenue with 1 million new projects a week, 9 June 2026. Source for the $500 million figure as Lovable stated it. https://techcrunch.com/2026/06/09/lovable-says-it-has-hit-500m-in-annualized-revenue-with-1-million-new-projects-a-week/\n\n[14] Sacra, Supabase at $170M/year growing 221% YoY, 11 June 2026. Source for Sacra's estimate of Supabase's ARR and growth; the $500 million raise is from TechCrunch, 5 June 2026. https://sacra.com/research/supabase-170m-year-growing-221-yoy/\n\n[15] Harvey, post on X, 1 October 2026. Source for the $400 million ARR and 97 per cent gross retention figures as Harvey reported them. https://x.com/harvey/status/2105675947693183113\n\n[16] Growth Unhinged, Kyle Poyar, The state of B2B SaaS and AI monetization in 2026, 13 May 2026. Source for the hybrid share, investor preferences and the AI credits figures. https://www.growthunhinged.com/p/the-state-of-b2b-monetization-in-2026\n\n[17] SaaStr, ten metrics from ICONIQ's 2026 State of AI report, July 2026. Source for what SaaStr reported of ICONIQ's consumption, outcome and subscription shares. https://www.saastr.com/the-builders-economy-10-metrics-from-iconiqs-newest-2026-state-of-ai-report/\n\n[18] Benchmarkit, 2026 SaaS and AI-native metrics benchmarks. Source for median net revenue retention by pricing model and the cohort size caveat. https://www.benchmarkit.ai/2026-saas-ai-native-metrics\n\n[19] Replit, Effort-based pricing, 18 June 2025. Source for the move to effort-based agent pricing. https://replit.com/blog/effort-based-pricing\n\n[20] Cursor, Teams pricing, June 2026. Source for the seat tiers with two usage pools. https://cursor.com/blog/teams-pricing-june-2026\n\n[21] Bessemer Venture Partners, The AI pricing and monetization playbook, 10 February 2026. Source for the gross margin ranges for AI companies and SaaS. https://www.bvp.com/atlas/the-ai-pricing-and-monetization-playbook\n\n[22] TechCrunch, Stripe didn't really buy OpenRouter because of the singularity, 19 August 2026. Source for what TechCrunch reported of the acquisition, its price and OpenRouter's take rate. https://techcrunch.com/2026/08/19/stripe-didnt-really-buy-openrouter-because-of-the-singularity/\n\n[23] TechCrunch, Anthropic acquires Vercept, 25 February 2026. Source for the acquisition and the shutdown of the Vy agent. https://techcrunch.com/2026/02/25/anthropic-acquires-vercept-ai-startup-agents-computer-use-founders-investors/\n\n[24] TechCrunch, Anthropic has acquired the dev tools startup used by OpenAI, Google, and Cloudflare, 18 May 2026. Source for the acquisition and the wind-down of all hosted Stainless products. https://techcrunch.com/2026/05/18/anthropic-has-acquired-the-dev-tools-startup-used-by-openai-google-and-cloudflare/\n\n[25] Mendral, Mendral team joins Anthropic, 21 July 2026. Source for the acqui-hire and the founders' statement about model progress. https://mendral.com/blog/mendral-team-joins-anthropic\n\n[26] SiliconANGLE, OpenAI acquires Ona, 11 June 2026. Source for the acquisition and the shutdown of Ona Cloud. https://siliconangle.com/2026/06/11/openai-acquires-ai-agent-orchestration-startup-ona/\n\n[27] TechCrunch, Meta acquires AI coding startup Conductor, 4 November 2026. Source for the acquisition and the cloud shutdown date. https://techcrunch.com/2026/11/04/meta-acquires-ai-coding-startup-conductor/\n\n[28] Google, Jules changelog, Jules turndown, 29 October 2026. Source for Jules turning read-only on 11 November 2026 and winding down on 31 December 2026. https://jules.google/docs/changelog/\n\n[29] Anchor Terminal, Sunsets, 10 November 2026. Source for the PlayHT, Hume and Runloop shutdown notices and the count of shutdown notices in the catalogue since November 2025. https://www.anchorterminal.com/sunsets/\n\n[30] CNBC, Salesforce to acquire Fin, 15 June 2026. Source for what CNBC reported of the price. https://www.cnbc.com/2026/06/15/salesforce-ai-customer-service-fin-acquistion.html\n\n[31] SiliconANGLE, Salesforce to buy Cresta for a reported $2.7 billion, 3 November 2026. Source for what SiliconANGLE reported of the price. https://siliconangle.com/2026/11/03/salesforce-to-buy-customer-service-automation-startup-cresta-for-reported-2-7b/\n\n[32] Dynatrace, Dynatrace completes acquisition of Arize, 1 October 2026. Source for the $915 million price and the completion. https://dynatrace.com/news/press-release/dynatrace-completes-acquisition-of-arize\n\n[33] SiliconANGLE, Nebius buys agentic search startup Tavily, 10 February 2026. Source for the price and earn-out as reported. https://siliconangle.com/2026/02/10/ai-infrastructure-giant-nebius-buys-agentic-search-startup-tavily/\n\n[34] OpenAI, API deprecations. Source for the Assistants API shutdown on 26 August 2026 and the Agent Builder shutdown on 30 November 2026. https://developers.openai.com/api/docs/deprecations\n\n[35] Android Authority, OpenAI is sunsetting ChatGPT Atlas, July 2026. Source for what Android Authority reported of the announcement and shutdown dates. https://www.androidauthority.com/openai-sunsetting-chatgpt-atlas-3686001/\n\n[36] CNBC, Anthropic's new tools spark software stock selloff, 6 February 2026. Source for what CNBC reported of the trigger and the moves. https://www.cnbc.com/2026/02/06/ai-anthropic-tools-saas-software-stocks-selloff.html\n\n[37] ETF Trends, Tech ETFs in Q1, the fracturing of one tech trade, 31 March 2026. Source for the IGV decline of 24 per cent in the first quarter, its worst since 2008. https://www.etftrends.com/tech-etfs-q1-fracturing-one-tech-trade/\n\n[38] CNBC, Software stocks plunge on ServiceNow and IBM results, 23 April 2026. Source for what CNBC reported of ServiceNow's record fall and Workday's year to date decline. https://www.cnbc.com/2026/04/23/software-stocks-plunge-on-servicenow-ibm-results-ai-fears-escalate.html\n\n[39] Menlo Ventures, 2026, the state of enterprise AI, 3 November 2026. Source for code-sandbox and search adoption among the AI builders surveyed. https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/\n\n[40] Sacra, Retell AI company profile, August 2026. Source for Sacra's estimate of Retell's annualised revenue and capital raised. https://sacra.com/c/retell-ai/\n\n[41] Bessemer Venture Partners, Cloud built the foundation, AI is now rewriting the rules, November 2025. Source for Fin's per-resolution price and the ARR trajectory Bessemer reported. https://www.bvp.com/atlas/cloud-built-the-foundation-ai-is-now-rewriting-the-rules\n\n[42] Chroma, Chroma Cloud wind-down. Source for the wind-down dates after the Replit acquisition. https://www.trychroma.com/cloud-wind-down\n\n[43] FlowiseAI, Flowise end of life discussion, 29 July 2026. Source for the code freeze, archive date and the maintainer's reason. https://github.com/FlowiseAI/Flowise/discussions/6727\n\n[44] OpenAI, OpenAI to acquire Promptfoo, 9 March 2026. Source for Promptfoo staying open source and supported after the acquisition. https://openai.com/index/openai-to-acquire-promptfoo/\n\n[45] Inworld, Migrate from PlayHT, 2026. Source for the PlayHT platform closure after Meta hired the team. https://inworld.ai/resources/migrate-from-playht\n\n[46] Bloomberg, Get me out, traders dump software stocks as AI fears erupt, 3 February 2026. Source for the start of the selloff on 3 February and the Jefferies trader who coined SaaSpocalypse. https://www.bloomberg.com/news/videos/2026-02-03/traders-dump-software-stocks-as-ai-fears-erupt-video\n\n[47] Salesforce, Salesforce completes acquisition of Fin, 10 September 2026. Source for the completion date and Fin being the company formerly called Intercom. https://www.salesforce.com/news/press-releases/2026/09/10/salesforce-completes-acquisition-of-fin/\n\n[48] Fin, Pricing, read 10 November 2026. Source for the $0.99 per outcome price. https://fin.ai/pricing\n\n[49] Menlo Ventures, Stripe to acquire OpenRouter, why everyone is obsessed with model routing, 19 August 2026. Source for Menlo's account of OpenRouter's uptime and cross-provider contracts. https://menlovc.com/perspective/stripe-to-acquire-openrouter-why-everyone-is-obsessed-with-model-routing/\n\n[50] InfoWorld, Cursor customers will lose access to OpenAI coding models in November, 31 August 2026. Source for OpenAI's decision and the 12 November shutoff. https://www.infoworld.com/article/4216503/cursor-customers-will-lose-access-to-openai-coding-models-in-november.html\n\n[51] Dealroom, Mercor doubles to $2B gross revenue run rate as AI labs buy expert data, 6 July 2026. Source for Mercor's stated gross run-rate and the share from OpenAI and Anthropic as reported. https://dealroom.co/news/137121-mercor-doubles-to-2b-gross-revenue-run-rate-as-ai-labs-buy-expert-data/\n\n[52] MedCity News, Anthropic says it is critical to bring company products to the EHR, February 2026. Source for the quoted remark about Abridge. https://medcitynews.com/2026/02/anthropic-its-critical-to-bring-company-products-to-ehr/\n\n[53] CNBC, Anthropic goes from software foe to friend, these stocks should benefit, 25 February 2026. Source for the analysts' list of data owners expected to benefit. https://www.cnbc.com/2026/02/25/anthropic-goes-from-software-foe-to-friend-these-stocks-should-benefit.html\n\n[54] ElevenLabs, ElevenLabs crosses $500M ARR and welcomes new investors, May 2026. Source for the $500 million ARR figure as ElevenLabs stated it. https://elevenlabs.io/blog/500m-arr-and-new-investors\n\n[55] TechCrunch, Anthropic co-founder on cutting access to Windsurf, 5 June 2025. Source for Anthropic ending Windsurf's direct access during OpenAI's bid. https://techcrunch.com/2025/06/05/anthropic-co-founder-on-cutting-access-to-windsurf-it-would-be-odd-for-us-to-sell-claude-to-openai\n\n[56] a16z, Joe Schmidt, Avoiding death on the Yellow Brick Road, 27 May 2026. Source for the Yellow Brick Road framing of horizontal tasks the labs win. https://a16z.com/avoiding-death-on-the-yellow-brick-road/\n\n[57] Anthropic, Claude Code sandboxing documentation, read 10 November 2026. Source for the open-sourced sandbox runtime. https://code.claude.com/docs/en/sandboxing\n\n[58] Colossus, Sarah's wager, July 2026. Source for Sarah Guo's quoted remarks on lab priorities. https://colossus.com/article/sarah-guo-conviction/\n\n[59] The Next Web, AI model costs are pushing startups towards cheaper open weight models, 21 September 2026. Source for what Bloomberg reported of Harvey's gross margin and token growth. https://thenextweb.com/news/model-costs-startups-open-weights\n\n[60] Harvey, Harvey raises $550M at a $15.5B valuation to help legal teams own their intelligence, 9 September 2026. Source for the round. https://www.harvey.ai/blog/harvey-raises-dollar550m-at-a-dollar155b-valuation-to-help-legal-teams-own-their-intelligence\n\n[61] Harvey, Making Harvey the orchestration layer for legal work, 15 October 2026. Source for the orchestration and connector positioning. https://www.harvey.ai/blog/legal-work-orchestration\n\n[62] SaaStr, Harvey puts a former practicing lawyer in every deployment, 2 September 2026. Source for the forward-deployed pod model as Harvey's product chief described it. https://cloud.substack.com/p/harvey-puts-a-former-practicing-lawyer\n\n[63] Ona, Anthropic model access in Ona, 5 August 2026. Source for the Ona Cloud shutdown on 31 October 2026 as Ona announced it. https://ona.com/stories/anthropic-model-access-in-ona\n\n[64] The Next Web, Lovable says two-thirds of the Fortune 500 use it and revenue hit $600m, 24 September 2026. Source for the $600 million figure as Lovable's CTO stated it. https://thenextweb.com/news/lovable-fabian-hedin-fortune-500-600m-revenue-humanx\n\n[65] OpenRouter, Pricing, read 10 November 2026. Source for the 5.5 per cent fee on credit purchases and no markup on inference. https://openrouter.ai/pricing\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-11",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.7",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "From the data",
        "url": "https://www.anchorterminal.com/blog/from-the-data/"
      },
      {
        "name": "Who gets paid",
        "url": ""
      }
    ],
    "description": "Quynh Tran-Thanh on how the 1,644 listings on Anchor Terminal charge, which models the top-graded tools use, where funding went in the last twelve months, what a year of shutdowns and acquisitions says about whether SaaS survives and whether the labs are eating the market, and which businesses thrive next to a lab.",
    "facts": [
      "Quynh Tran-Thanh",
      "10 November 2026"
    ],
    "h1": "Who gets paid in the agent stack. Seats, meters and the year the labs went shopping",
    "image": "https://www.anchorterminal.com/assets/og/blog-who-gets-paid-in-the-agent-stack.png",
    "path": "/blog/who-gets-paid-in-the-agent-stack",
    "published": "2026-11-10",
    "section": "blog",
    "title": "How the agent stack charges in 2026, and who got bought | Anchor Terminal",
    "toc": null,
    "updated": "2026-11-10",
    "url": "https://www.anchorterminal.com/blog/who-gets-paid-in-the-agent-stack"
  },
  "tokens": {
    "markdown": 9450,
    "slim": 5080
  },
  "version": 1
}