Fentner

Markets · Opinion

Stop paying per seat for software that is mostly someone else's model

If SaaS products are turning into thin harnesses around rented models, buyers should split the bill into harness fee and model cost, and stop signing long seat deals.

The most celebrated AI application in software reportedly loses money on every dollar it books. According to CloudZero's account of reporting by The Information, Cursor ran a gross margin of negative 23% in the quarter ended January 2026, while approaching $2 billion in annualised revenue. One commentator worked that through to $1.23 paid to Anthropic and OpenAI in API costs for every dollar Cursor took in. If the vendor at the centre of the AI coding boom cannot price its product above what the model costs it, I see no reason to trust the seat price on any other AI-heavy invoice in your stack.

A post on blog.sshh.io this week argues that every SaaS business is becoming a harness: context, integrations, permissions and review screens wrapped around a stateless model. I agree with the diagnosis. I expected the author to follow it to pricing, and instead he follows it to the org chart. The buyer's conclusion is simpler. If the product is model output wrapped in a vendor's plumbing, then a fee per human login is the wrong unit. You are paying per person for work increasingly done by agents, and the real cost of that work sits with a model lab your vendor rents from.

Investors have already started doing this arithmetic. On 3 February 2026 Anthropic pushed eleven plugins for Claude Cowork to GitHub, and by one analysis $285 billion came off software stocks in a single day, with losses reaching $1 trillion by mid-month. CNBC reported that the S&P 500 Software & Services Index, 140 companies, fell more than 4% that Thursday, its eighth straight losing session, leaving it down about 20% for the year. Thomson Reuters, Salesforce and LegalZoom were among the hardest hit. Constellation Research told CNBC the sell-off reflected worry about how much software companies can charge. I think that reading is right, and procurement teams should act on it before vendors' finance teams do.

Vendors will act, and not in your favour. Newcomer reported in 2025, citing sources familiar with the figures, that Cursor was running up large Anthropic bills and scrambling to plug its pricing leaks. A subsidised seat ends in one of two ways: a price rise at renewal, or a usage cap that quietly shrinks what the seat you already paid for will do. A buyer locked into a three-year per-seat contract carries exactly the risk that the vendor's margin cannot.

The vendors' reply is that the harness is where the value lives, and Cursor appears to prove it by getting thicker. Following the SpaceX arrangement, reported on X as an option to buy Cursor for $60 billion, Cursor's plans now include generous usage of Grok 4.6, Grok 4.5 and Composer 2.5 as first-party models. A vendor that owns its model, the argument goes, can charge a flat seat with a clear conscience. Look at what Cursor actually did, though. Its own models are bundled, and third-party models from Anthropic, OpenAI and Google are billed at each model's own rates. Cursor split the bill into harness and pass-through model cost. Its margins forced it there. I would ask every other vendor for the same split now, before their own model invoices force a cruder version on you.

Intercom shows the other route, and its trap. Fin is reportedly nearing $100 million in ARR, roughly half of Intercom's revenue, at $0.99 per resolved conversation, after a $49 monthly base that covers 50 outcomes. That is a price tied to work done, which is what buyers should want. But one pricing index describes Fin as outcome-based and optionally stacked on seats. I suspect this is the future most incumbents would like: keep the seat, add the meter. If an agent resolves the ticket, the resolution is what you are buying, and paying again for a human seat that no longer touches that ticket is paying twice for one piece of work.

The harness post makes one more point buyers should take seriously. Its author says companies should own the top-level harness themselves and plug vendor products into it for specific workflows, swapping them out as third parties get better. A component you plan to swap has no business sitting on a multi-year contract sized to a headcount you are planning to shrink.

Before any renewal this year, ask the vendor to break its quote into model cost and harness fee, name the models it runs, and put in writing what happens to your price if its model bill rises. A vendor that cannot answer is either losing money on you, as Cursor reportedly was, or hiding a margin it would rather you not see. In both cases, sign for twelve months and count only the seats of people who will still be logging in next year.

Prompted by The Harness Is the Company, blog.sshh.io.