AI killed the free tier. Agents are killing subscriptions.
The Free Tier
Software wasn't always cheap to run. In the on-premises era, every deployment was bespoke: license fees, integrators, switching costs measured in years. Siebel, the dominant CRM before Salesforce, averaged $6.59 million per implementation over three years.[1] 84% of the buyer's spend went to integrators and hardware, not to the software vendor. The vendor's gross margin on the software license itself was 97%.[2] Delivering the software wasn't the expensive part. Deploying it was.
Cloud eliminated the expensive part. Multi-tenant architectures let one codebase serve every customer. The deployment cost that integrators had captured disappeared. Dropbox Plus costs $11.99 per user per month. Adding one customer to the platform costs roughly $2.84.[3] That's 24% of one seat's revenue. At 10 seats, it's 2.4% of revenue. At 100 seats, it's a rounding error. SaaS gross margins settled at 75–85%: Atlassian at 83%, Dropbox at 80%, Adobe at 88%. Ben Thompson's Aggregation Theory captured the logic: "Zero distribution costs. Zero marginal costs. Zero transactions. This is what the Internet enables."
Incumbents like Salesforce used those margins to entrench. 40–45% of revenue went to sales and marketing, building a customer base that was almost impossible to leave. Migrating off Salesforce takes 6–18 months. Five years of history doesn't export cleanly. Once customers were in, they stayed.
Competitors couldn't match that spend. They didn't have the revenue. So they found a different weapon: the free tier. If marginal costs are near zero, give the product away and let it spread. Dropbox reached 700 million users, 18 million paying, $2.55 billion in revenue on a 2.6% conversion rate. The math works because the other 97.4% cost almost nothing to serve. HubSpot, Slack, Figma, Notion, Airtable, Zapier, Canva — product-led growth became the challenger playbook. According to OpenView, who coined the term, PLG companies grew at roughly twice the rate of sales-led peers.[4]
But giving the product away to hundreds of millions of users while waiting for 2–3% to convert takes years of runway. A decade of cheap capital provided it. Hundreds of companies entered - and because switching costs kept existing customers in place and the market kept producing new ones, margins never compressed. High margins, growing demand, cheap money. In 2021, investors priced the combination at 24× revenue.
AI Killed the Free Tier
By late 2022, it was 6×. Rising interest rates ended cheap capital. The post-COVID demand surge that had absorbed hundreds of new entrants was fading, and those companies were now competing for the same customers. Salesforce dropped 59% from its peak.[5] The growth-over-profit narrative that had justified premium multiples for a decade stopped working. The industry needed a new story.
It found one in ChatGPT. Within months of its November 2022 launch, every major SaaS company was adding AI features. If you're sitting on years of customer data and 6–18 month switching costs, AI doesn't require a land grab. You upsell the base you already have. AI features trained on your customers' data are immediately useful, and they deepen the moat: switching now means losing not just your workflows, but the intelligence built on top of them. Salesforce launched Einstein GPT. Microsoft embedded Copilot across Office.
Incumbents weren't the only ones who saw the opening. Anthropic launched Claude in March 2023. Perplexity raised a $25 million Series A the same month. LLM wrappers flooded the market. Y Combinator's batches went from 15% AI to 40% in a single year. SaaS stocks recovered 58% on the year.[6] Once again, demand was outpacing competition.
But AI giveth, and AI taketh away. Zero marginal costs. Every time a user triggers an AI feature - Notion's AI assistant, GitHub Copilot's code suggestions, Canva's Magic Studio - the company incurs real inference costs. The near-zero marginal cost that the entire model depended on started eroding. The result is a prisoner's dilemma. You need this magical new AI to compete. You need to stay free to preserve product-led growth. You have to do both. You can't afford to do both. You try anyway.
Microsoft was losing $20 per user per month on GitHub Copilot, with heavy users costing $80, while charging $10. GitHub had 100 million developers and an opportunity to put AI in front of every one of them.[8] Cursor hit $1 billion ARR at negative 30% gross margins, a deliberate bet on a new category.[9] Notion doubled its cost basis from 10% of revenue to 20% in two years, bolting AI onto your team's tribal knowledge.[10] The bet is lock-in. Absorb the losses, remain irreplaceable (or become irreplaceable), raise prices later.
Can Ads Save the Free Tier?
The internet has a standard hedge for this sort of bet. Google Search earns roughly 4.5 cents per query against an operating cost of about 2.7 cents, yielding 41% operating margins across $225 billion in 2025 search revenue.[7] Advertising is the most proven mechanism for making digital services universally accessible. The math works for AI too: a typical query costs $0.003–$0.01, and a contextual ad impression at search-level CPMs generates $0.04-$0.06. The ad covers the query with margin to spare.
At ZeroClick, we recognized AI would need this. We started as Pie in 2024, building for the open web with a conviction that advertising doesn't have to be adversarial, that you can align the incentives of users, developers, and advertisers so everyone benefits. AI made that conviction urgent. Our hypothesis wasn't that we'd do ads better inside walled gardens. It was that AI wouldn't be just ChatGPT — that millions of AI developers and thousands of AI platforms would emerge outside the walled gardens, and they'd need monetization infrastructure. We launched ZeroClick in August 2025, raised $55 million, and built a platform that gives those developers full control over how and when monetization appears in their products.[11] AI platforms get funded. Users get free access. Advertisers reach high-intent audiences.
Models have real limitations that good commercial context can address. Training data has cutoffs. Brute-force web search burns tokens inefficiently. Curated, current advertiser context — a caching guide from Redis, a deployment tutorial from Netlify - can genuinely make an AI's answer better. We're working on deeper integrations that go beyond answers that have even more utility. Our model works. We've proved that ads can extend free access for query-scale interactions between humans and AI.
But AI usage is exploding far past the bottleneck of interactions between humans and AI.
Agents Are Killing Subscriptions
Anthropic's engineering team spent $20,000 building a C compiler from scratch with Claude Opus 4.6. Two billion tokens over two weeks.[12] OpenAI's ChatGPT Go plan costs $8/month. That compiler would require more than 200 years of Go revenue. And the range keeps expanding. Context windows are growing into the millions of tokens. METR estimates autonomous task horizons reaching 14.5 hours, with a confidence interval spanning up to 98 hours.[13] Agents are running longer, operating with less human oversight, consuming more compute per task. A subscription is a fixed number. The costs are not.
Anthropic's Claude Code Max costs $200/month. Forbes reported on March 5 that the plan can consume roughly $5,000 in compute.[14] Anthropic can afford that ratio, for now. It raised $30 billion the same month at a $380 billion valuation.[15] But even Anthropic has limits. When OpenClaw, the most popular autonomous agent platform on OpenRouter, consumed 11.1 trillion tokens, Anthropic and Google cut off its access to their subscription-tier products. The usage-based APIs remain available. The subscriptions don't.
The obvious counterargument: costs are falling. Epoch AI estimates inference costs are halving roughly every two months, dropping around 40× per year.[16] But Jevons paradox is once again in vogue: cheaper tokens don't reduce total cost. They enable longer, more autonomous workflows that consume the savings and then some. Per-token costs are falling. Per-task costs are not.
The real counterargument is simpler: it's worth it. Even at $5,000 in actual compute, a Claude Code Max subscription costs less than half an entry-level software engineer. The question is how to charge for it. Low enough to lock customers in. High enough to survive the costs.
The Repricing
On January 14, Stripe completed its $1 billion acquisition of Metronome, rebuilding billing infrastructure for a world that no longer prices by seat.[17] Across the industry, 68% of SaaS vendors had already restricted AI features to premium tiers.[18] A CFO at a data infrastructure company captured the tension: "We're not monetizing AI to juice revenue. We're monetizing to avoid eating $10K of costs on a $500 plan."[19]
On February 3, the market delivered its own repricing. Roughly $285 billion in software market cap evaporated in a single day. Traders at Jefferies coined the term SaaSpocalypse. Over the following weeks, the sector lost approximately $2 trillion.[20] HubSpot fell 62% from its peak. Workday's CEO stepped down. SAP dropped 14% despite 26% cloud revenue growth. By February 6, the median public SaaS company was trading at 3.6× revenue. In 2021 it was 24×.
The crash reflected what was already happening on the ground. Retool reported on February 17 that 35% of companies had already replaced a SaaS tool with a custom build.[21] Garry Tan, CEO of Y Combinator: "Why pay $30/seat/month for overbundled SaaS when even non-tech ops people can vibe-code a custom solution in a weekend?"
Companies repriced. Notion launched Custom Agents on February 24 with credit-based pricing. Users found they cost 50–100× more than the same task via direct API. They knew this because Notion's own MCP server already lets Claude Code read and write to Notion directly. Clay repriced on March 11, splitting into data credits and a new metric called Actions.[22] An enrichment that costs $1.86 per lead through Clay costs $0.15 at direct API rates. But Clay has already taught its users to think in APIs, and they've already started to call them directly.
If you think these new buyers are tough - claude-code-pilled, X-scrolling, building over buying - just wait.
What Comes Next
The buyer in the agentic IDE building with one agent is now orchestrating a fleet of 100. In fact, the agentic IDE is now a cloud agent orchestration platform.[23] On OpenRouter, the largest open model router, the most popular apps are no longer agentic coding tools for humans. They are autonomous agents. OpenClaw leads the way at 11.1 trillion tokens in the last month. Daytona pivoted its entire company to agent infrastructure. CEO Ivan Burazin: "the market for serving AI agents is orders of magnitude larger than serving all human developers."[24] Aaron Levie, CEO of Box for twenty years, argued on March 4 that enterprises will soon run "100× or 1,000× more agents than people."[25] Agents don't attend webinars, respond to brand campaigns, or evaluate products through demo calls. They read API documentation, test integrations programmatically, and select on functionality and cost. Or at least they should.
The infrastructure for this shift is still being built. Most developer tools don't support account creation via API.[26] Even when agents can access services, they can't verify what they're getting. Cisco found vulnerabilities in 26% of agentic skills on OpenClaw - prompt manipulation, data exfiltration, capability escalation.[27] Stripe built an Agentic Commerce Protocol for settlement: disputes, refunds, liability.[28] On March 18th it announced the Machine Payments Protocol, an open standard for agents to transact autonomously, but nothing yet covers discovery or selection. When an agent chooses a service, no standard mechanism confirms whether that choice was quality-driven or manipulated. This is Jevons paradox again. Models reduce entropy. They organize, predict, retrieve. But the agents and services they enable create new entropy - faster and of greater magnitude - than the models resolved.
Now, agents want to organize more than software. Sequoia's Julien Bek published "Services: The New Software" on March 5.[29] For every dollar spent on software, six are spent on professional services — accounting, legal, recruiting, consulting. If agents can perform the work, not just support it, those budgets are addressable. Anthropic's Code Review already prices this way: $15–25 per pull request, charged against the cost of a shipped bug, not against a seat.[30]
Code Review is one service offered by one agent harness built by one company. It is a particularly massive service, from a generational company, and it will still be hotly contested. There are millions of other services. And there will be trillions of other agents. These agents and their companies will have unique data and unique capabilities that aren't in model training data. That will make them valuable. It will also make them hard to find, hard to compare, and hard to trust. Humans have search, ads, reviews, testimonials, verifications, marketplaces - infrastructure that reduces the entropy of commerce. Agents will need their own.
Sources
[1] Nucleus Research, "Real ROI from Siebel Reference Customers" (2002)
[2] Siebel Systems 10-Q, Q3 2004 (SEC EDGAR)
[3] CloudZero, analysis of production SaaS environments
[4] OpenView Partners, PLG benchmark data from 450+ companies
[5] MacroTrends, Salesforce (CRM) historical stock price — $309.96 (November 8, 2021) to $128.27 (December 16, 2022)
[6] iShares Expanded Tech-Software Sector ETF (IGV), +58.47% calendar year 2023 (Morningstar)
[7] Alphabet FY2025 earnings (February 4, 2026); Google Blog, "more than 5 trillion searches per year" (March 4, 2025)
[8] GitHub, "100 million developers" (January 2023); Wall Street Journal, Microsoft Copilot unit economics
[9] Newcomer (Tom Dotan, August 2025) and Foundamental analysis — Cursor gross margins
[10] Wall Street Journal, Notion CEO Ivan Zhao on AI infrastructure costs
[11] ZeroClick, "ZeroClick launches with $55 million to build the ad network for AI" (August 20, 2025)
[12] Anthropic Engineering, "Building a C compiler with a team of parallel Claudes"
[13] METR, "Task-Completion Time Horizons of Frontier AI Models" (March 3, 2026)
[14] Forbes, Claude Code Max compute costs (March 5, 2026)
[15] Anthropic, "$30 billion Series G funding" (March 2026)
[16] Epoch AI, "LLM Inference Price Trends" (updated February 2026)
[17] Stripe, completion of Metronome acquisition (January 14, 2026); Lago, "Why Stripe paid $1B for Metronome" (February 14, 2026)
[18] BetterCloud, "2026 SaaS Industry Report"
[19] Metronome, "2025 AI Pricing Field Report"
[20] Fortune, "Why that $2 trillion software wipeout didn't derail the AI bull market" (February 10, 2026)
[21] Retool, "2026 Build vs. Buy Report" (February 17, 2026)
[22] Clay, "Introducing Clay's New Pricing" (March 11, 2026)
[23] Cursor, "The Third Era" — cursor.com/blog/third-era
[24] Ivan Burazin (Daytona CEO), on agent infrastructure pivot (March 2026)
[25] Aaron Levie, CNBC interview (March 4, 2026) and Latent Space podcast
[26] Jared Friedman (Y Combinator), on API-first for the agent era (February 27, 2026)
[27] Cisco Blogs, "Personal AI Agents like OpenClaw Are a Security Nightmare" (January 28, 2026)
[28] Stripe Docs, "Agentic Commerce Protocol"
[29] Julien Bek (Sequoia Capital), "Services: The New Software" (March 5, 2026)
[30] Anthropic, "Code Review" (launched March 9, 2026)