Seats Are Dying. What the 2026 SaaS Repricing Looks Like From the Compute Layer

On 3 February 2026, Thomson Reuters had its worst trading day on record, down 18%. RELX fell 14%, its worst since 1988. Gartner dropped 21%. The trigger, per Reuters, was that Anthropic had released open-source plugins for its Cowork product covering legal, financial and data-analysis work the previous Friday. Not a product launch. Plugins.

I run infrastructure that AI agents execute code on, and I sell it by the second. So I watched that week with unusual interest, because what the market was repricing was not really software companies. It was the unit of software. For thirty years the unit was a seat: a human, sitting at a screen, paying monthly. In 2026 the unit is becoming an outcome, an action, a resolution, or a token, and every one of those is a metering problem that lands on people like me.

This post is what that shift looks like from underneath: what vendors are actually charging now, what it did to the companies that grew fastest, what it costs to run, and what broke. Every number links to its source. Where I only have secondary reporting I say so.

The repricing, in one table

Satya Nadella said the quiet part in December 2024, on the BG2 podcast: business applications "are essentially CRUD databases with a bunch of business logic", and "the business logic is all going to these AI agents" (transcript). Fourteen months later the vendors had rewritten their price lists to match.

Vendor	Old unit	New unit	Price	Source
Salesforce Agentforce	Seat	Conversation, then "Flex Credit" action	$2 per conversation; $0.10 per action (20 credits at $500 per 100,000)	Salesforce Help
Microsoft Agent 365	Seat	Human who "manages or sponsors agents"	$15 per user per month; GA 1 May 2026	Microsoft
Zendesk	Seat	Automated resolution (ticket stays closed 72 hours)	$1.50 committed	Zendesk
Fin (ex-Intercom)	Seat	Outcome: resolution or procedure hand-off	$0.99 per outcome; $49 base with 50 included	Fin
ServiceNow	Seat	"Assists" pooled per seat; small agentic action 25, large 150	Tiers effective April 2026 (per partner documentation)	Crossfuze

Three observations about that table.

Microsoft's is the cleverest. Agent 365 does not charge for agents; it charges for every human who touches one. It preserves the seat by redefining what sits in it. Whether that survives contact with a customer who runs a thousand agents supervised by three people is the question Microsoft will answer in its 2027 renewals.

Zendesk's 72-hour rule is the one to watch, because it is the first pricing term that puts the vendor's revenue at risk if the AI is wrong. A resolution that reopens on day two is free. I expect that clause to be in every support contract within eighteen months, and I expect a lot of engineering effort to go into making tickets stay closed.

And Salesforce spent roughly $3.6 billion in June to acquire Fin, the company charging $0.99 per outcome with a claimed 76% end-to-end resolution rate. When the incumbent that invented the seat buys the company that invented the resolution, the argument is over.

The aggregate data is thinner than I would like. The most credible survey I found is Kyle Poyar's 2026 B2B monetisation study of 230-plus companies, reported second-hand: 38% now use usage as the primary model, up from 27% in 2023, 61% run a hybrid, and only 8% are purely per-seat. I could not verify that against the primary report and would not build a slide on it. The direction is not in doubt.

The companies that grew fastest have no seats at all

The other side of the repricing is the app-generation companies, and their numbers are the most extraordinary in software this decade.

Lovable went from launch to $100 million in annualised revenue in eight months, raised at $6.6 billion in December, and last month confirmed a $13.3 billion valuation on a $500 million run-rate reached in June, with 60 million projects created. Replit went from $3 billion to $9 billion in six months and has said publicly it is targeting $1 billion in ARR by year end. Emergent hit $100 million ARR eight months after launch. Wix bought Base44 for $80 million up front in June 2025 when it was doing "a few million" in ARR, and its CEO said in August it would be at $40 to $50 million by year end. And Cursor, the coding agent, was acquired by SpaceX in an all-stock deal valued at $60 billion that closed three weeks ago.

None of these companies sells a seat in the old sense. They sell credits, messages, tokens, or a subscription whose real limit is inference budget. That is the same unit as the enterprise vendors are moving to, arrived at from the opposite direction. The seat companies are adding usage because agents replaced the human. The agent companies never had a human to charge for.

What it costs to run, which is the part I actually know

Here is where I stop reporting and start speaking from experience, because I run the layer that this business model is built on.

Every one of those products has the same cost structure: a model call, then a sandbox in which the generated code runs, then hosting for whatever gets built. The first is paid to Anthropic, OpenAI or Google. The second and third are paid to someone like me, or built in-house.

The model prices as of this week, per million tokens input and output, from the Anthropic and OpenAI price pages: Claude Fable 5.1 at $10 and $50, Opus 5 at $5 and $25, Sonnet 5 at $2 and $10, GPT-6 Astra at $10 and $50, GPT-5.6 Terra at $2 and $12. Sonnet 5 is a third cheaper than Sonnet 4.6 was. Prices fall roughly annually; usage rises faster. Anthropic said in May it had reached a $30 billion run-rate, with Claude Code alone at $2.5 billion by February.

The margin consequence is brutal and public. The Information reported, per CloudZero's summary, that Cursor's gross margin was negative 23% in the quarter ending January 2026 at about $2 billion ARR, and that its in-house model was what turned enterprise margins positive. I cannot verify the number against the primary report, but it matches what I see: a company reselling frontier inference at a flat monthly price is running a hedge fund, not a software business, and the June 2025 Cursor pricing change that swapped 500 requests for "$20 of usage" (followed by an apology and refunds) was the hedge being unwound in public.

The sandbox layer is where I live, and 2026 is the year it became visible as an industry:

Provider	Isolation	Scale claim	Source
E2B	Firecracker microVMs	1B+ sandboxes launched; Perplexity, Manus, Hugging Face as customers	e2b.dev
Modal	Own runtime	$355M Series C at $4.65B in May; over $300M annualised revenue; 1B+ sandbox launches	Modal
Vercel Sandbox	Firecracker microVMs	25M sandboxes per week; 15M builds per day	Vercel
Cloudflare Sandboxes	One VM per container	GA April 2026; active-CPU billing; Figma as named user	Cloudflare
Daytona	Not stated	$24M Series A in February; LangChain, Writer as customers	PR Newswire
PandaStack (mine)	Firecracker microVMs	Open source; 179ms p50 create	pandastack.ai

Notice the billing models. Cloudflare bills active CPU, not wall-clock. Vercel's "Fluid" compute does the same. Modal always did. This is not a coincidence: when your customer is charging per outcome, they cannot afford to pay you for idle. I built scale-to-zero into my platform because I had to, and I wrote about what it cost; the short version is that every provider in that table has made the same trade for the same reason. The unit of software changed, so the unit of compute had to change with it.

The other thing the table tells you is that Firecracker won. Three of the six named providers, including the two largest by volume, run the same open-source hypervisor I do. When I chose it in early 2026 that was a bet. It no longer is.

What broke

Outcome-based software running on ephemeral compute has failure modes that seat-based software did not, and 2025 and 2026 produced a clean set of examples.

Vibe-coded apps ship without authentication. In May 2025 a Replit engineer disclosed CVE-2025-48757: Lovable-generated apps were routinely deployed with Supabase row-level security misconfigured, exposing user data to anyone with the public key. Secondary reports put it at 170 of 1,645 scanned apps. In April 2026 Lovable itself had a ten-week regression in which the chat history and source of public projects was readable by any logged-in user; they fixed it within two hours of public disclosure, but the first researcher report had arrived two months earlier. In May, a scan reported via Security Boulevard found more than 5,000 generated apps on Lovable, Replit, Base44 and Netlify with no authentication at all.

Agents delete production. In July 2025 a Replit agent deleted the production database of SaaStr during a declared code freeze, taking records for over 1,200 executives with it. Replit's CEO called it "unacceptable and should never be possible", and the fixes were the ones any infrastructure engineer would list: separate dev and production databases, one-click rollback, a planning-only mode. I have deleted my own production data by hand under similar confidence, so I am not going to be smug about it. The point is that the guardrails were not there because the product was not designed by people who had been paged.

The agent platform itself leaks. In February 2026 Wiz found Moltbook, a social network for AI agents, running Supabase with no row-level security and a hard-coded key, exposing 1.5 million agent tokens. Same root cause as the Lovable apps, one layer up.

The pattern across all of them: the outcome was priced and the infrastructure was not designed. When a resolution is worth $1.50 and a generated app is worth a $20 subscription, there is enormous pressure to make the compute behind it as cheap and as automatic as possible, and the first thing that gets automated away is the person who would have asked whether RLS was on.

What I think happens next

The seat is not going to zero. Microsoft's $15-per-human pricing is a bet that it survives as the billing anchor even when the work is done by agents, and Microsoft usually wins bets about billing anchors. But the seat is no longer where the growth is, and the companies with the steepest curves in software this year have never sold one.

For those of us in the compute layer the consequences are already here. Bill for active time, not allocated time, or lose the customer whose own revenue depends on that. Make creation cheap enough that a $0.99 outcome can afford a fresh microVM, which is why I measure my boot path in milliseconds and refuse to stop the clock early. And build the guardrails the product companies will not, because when the agent deletes the database, the customer is going to ask the infrastructure provider why that was possible.

Cloudflare's CEO said in June that bots now exceed human traffic on their network for the first time. That is the same shift, seen from the network instead of the invoice. The customer of software is increasingly a program. Programs do not sit in seats. They consume, they produce outcomes, and they expect the bill to say so.

---

Related: What Scale to Zero Cost Me to Build and AI Agent Cloud: What the Term Means.