The Cost of Free

The door turns easily on the way in. The cost is on the next page of the contract.

I pay Anthropic every month. Claude Code has changed how I work, and I would recommend it to any senior operator seeking to build a real technical advantage. I also know that I am locked in.

That admission is the right place to start, because every major AI vendor is running the same play, and pretending otherwise is the first mistake a thoughtful executive can make. The product is excellent. The pricing is generous. The lock-in is the strategy. Free is the wedge.

The pattern is a forecast

In January 2025, Microsoft made Copilot Chat free for any user with a work account. By December of the same year, it announced base price increases across Microsoft 365, effective July 1, 2026. The chat was the wedge. The renewal was the trade.

Salesforce launched Agentforce in September 2024 at $2 per conversation. Eight months later, it introduced Flex Credits at $0.10 per action. Six months later, a per-user add-on at $125 per seat per month and an Editions tier starting at $550 per seat per month were introduced. Three pricing models in 18 months. The repricing was not indecision. Conversation pricing had already produced a telling result: Salesforce signed 5,000 Agentforce deals in the first two quarters, and only 3,000 paid. Many buyers could not model the cost of a single query that triggered eight backend processes and still counted as one $2 conversation. Opaque pricing works precisely because the buyer is too tired to model it.

Slack retired its $10 AI add-on on August 17, 2025. Anyone who had bought it kept it until their next renewal, then moved to the Business Plus tier at $15 per user per month, or $18 per user per month billed monthly. Atlassian made Rovo free within its Premium and Enterprise plans in April 2025, and Standard later that year, with a footnote that overages will become billable with at least 90 days' notice. ServiceNow's Now Assist passed $600 million in annual contract value in 2025, more than doubling year over year, on the way to a $1.5 billion AI target, with its pricier Plus-tier deals quadrupling. In each case, the AI is the lever, and the tier shift is the goal. They are building the dependency now and pricing the consumption later.

None of this is hidden. The disclosures live in earnings calls, on pricing pages, and in customer letters. Why it works anyway, and what to do about it, is the rest of this piece.

The moat moved to the wrapper

The Anthropic story is the cleanest illustration I have, because I am inside it.

The Claude stack is expensive, and Opus 4.7 in particular runs hot. Claude Code's agentic loops are the whole point of the tool. When I turned them on, the background cycles I call the Paperclip Heartbeats started firing, and my monthly bill jumped. The agent does not sleep. Every wake cycle consumes a context window, every context window is a transaction, and every transaction runs through one provider's meter. I'm not complaining; I made the choice. I am naming the mechanism.

So I went looking for the exit, and I found Aider. Aider is what an honest software architect would build. It separates the model from execution by design. It is model-agnostic, runs against any provider's API, and costs only what the underlying API costs. The architecture is cleaner than Claude Code's and more honest.

I went back to Claude Code within a week.

The reason is what gets built around the model. The plan mode, the stop hooks, the sub-agents, the Skills layer, the connected tools, the autonomous loops, all of it stitches into something that does not pull apart cleanly. Aider has the better separation of concerns. Claude Code has the better unified product. Hold on to that contrast. It is the whole argument.

This is the part most executives miss. The story since 2023 has been about model capability: which model is smarter, which model tops which benchmark. That story is two years out of date. For most enterprise work, the frontier models are close enough in raw capability that the gap rarely decides the outcome. The real differentiation is the wrapper, the layer built around the model. Claude Code's $2.5 billion annualized run rate, reached in February 2026 and more than doubled since the start of the year, measures the strength of that wrapper. The model is becoming a commodity. What surrounds it is the moat.

Why this lock-in cycle is different

A reader might reasonably ask, "So what?" Smart vendors build sticky products. That is capitalism doing what capitalism does; why this lock-in cycle differs from the ones before it is worth four short points.

First, the speed. Cloud lock-in compounded over a decade. AI lock-in compounds in quarters. Once your team builds workflows around plan mode or Cursor's Composer or Now Assist, the prompt libraries, the agent orchestration, and the team habits calcify within months.

Second, the gravity. AI tools accumulate context in ways that prior software did not. Fine-tuned models, organizational memory, custom tool configurations, the institutional voice baked into your prompts, none of it ports cleanly. Salesforce's Data Cloud is one gravity well. Epic's Cosmos dataset is another; its value compounds with the data already inside it. The skills and tool registries that sit under an agent pull the same way.

Third, the psychology. The zero-price effect, documented by Kristina Shampanier, Nina Mazar, and Dan Ariely in 2007, skips calculation and triggers a category shift. Two dollars is a price. Zero is a different mental object. The IKEA effect compounds it: once a team has built a prompt library or trained itself on one vendor's tools, it values that work disproportionately because it made it.

Fourth, the asymmetry of attention. The vendor's best people work on your retention full-time: the pricing team, the growth team, the customer success team, all of it. Your team works on portability when there is room in the quarter, which is never. Across a three-year arc, that asymmetry compounds the way interest compounds. The renewal price measures one thing: what their model says you can afford and still stay.

The Broadcom acquisition of VMware is the clearest evidence of the asymmetry, even though it predates the AI cycle. AT&T's email to Broadcom CEO Hock Tan, filed in the New York State Supreme Court, described a renewal offer that would have raised annual VMware costs by roughly 1,000%. AT&T had been planning a normal renewal. Broadcom had been preparing the repricing for much longer and held all the power in the room. The 1,050 percent number is shocking. The deeper point is that the offer existed at all. Vendors model your dependency for years before they exercise it.

The move: rent the intelligence, own the execution

So here is the move, and it is the part most of these conversations never reach. Decouple the intelligence from the execution.

Two layers are being deliberately fused, and you should pull them apart. The intelligence is the model. It is commoditizing, it leapfrogs every few months, and it should be treated as a swappable input. The execution is everything around the model that actually does the work: how tasks are orchestrated, how the system reaches your data and tools, how it remembers, how you test and constrain its outputs, the governance around it, and the prompt and skill libraries your people accumulate. That layer is the durable asset. Done well, it is an operating system for agents, and any model plugs into it as a replaceable part.

The vendor's strategy is to weld those layers together, so that leaving the model means abandoning the platform your people built their days around. The executive counter-move is to keep the seam visible. Rent the intelligence. Own the execution. Build the layer that does not change as quickly as the models do; keep it portable; and let the model be the part you swap when a better one arrives, or a renewal turns hostile.

The same wrapper that traps you when it is rented becomes your compounding capability when you own it. Your prompts, your evaluation suites, your tool integrations, your governance, the institutional voice in your agents, all of it should outlive any single model and belong to you. The diagnostic is one question: of the execution you depend on, how much do you own, and where is the seam at which you could change the model? If there is no seam, you have not decoupled anything.

The reason this is more practical now than it was eighteen months ago is that the substrate is maturing. The frontier models have converged, making them substitutable. An open standard for connecting models to tools and data, the Model Context Protocol emerged in 2024 and was adopted by OpenAI, Google, Microsoft, and AWS, and later moved to a vendor-neutral body under the Linux Foundation in December 2025. Model gateways now allow a single system to route across providers. And recent lock-in shocks have quickly opened exit valves: when HashiCorp and Redis changed their licenses, the OpenTofu and Valkey forks followed, and the major clouds backed them. The tools for decoupling are no longer theoretical.

Be honest about the cost, because your CFO will be. Decoupling is not free either. Owning an execution layer is real engineering and governance work, and a fused product genuinely wins on speed and polish today, which is exactly why I use one. The judgment is where to fuse for velocity and where to decouple for durability, decided on purpose rather than by default, so that a convenience you adopted this quarter never quietly settles the question for you.

Five questions before you say yes

Five questions turn that judgment into a decision you can defend. Run them before saying yes to any free AI offer in the next twelve months, and run them again at every renewal.

The reversibility test. At the end of this contract, can you actually leave? Name the cost in dollars and in months. If you cannot name it, you do not know it.

The concentration test. What share of your stack runs through this vendor after this change? Count the AI features, the data layer, the integration surface, and the people who only know this tool.

The renewal test. Who holds the power at the next renewal, you or them? If the feature is free today, who controls its price three years from now?

The skills test. Are your people building portable skills or vendor-specific ones? When a developer learns one provider's plan mode, what fraction of that learning travels?

The optionality test. What futures are you closing off by accepting this offer? Name two. If you cannot name two, you have not thought about it.

These are calibration tools, and the decision still belongs to the executive who has to live with the renewal.

What free actually costs

The philosophical landing is the part most executives skip, and it is the only part that holds the rest together.

Free is a transfer of optionality, not money. Turning on free Copilot Chat did not save you anything on the invoice. It costs you something the invoice never shows: the future right to walk away from Microsoft at your next renegotiation. That option had value. You gave it up for immediate convenience, and the trade was complete the day you said yes.

The deeper point is about identity. Past a certain depth of integration, your workflow becomes your tool. Hospitals that have run Epic for fifteen years are no longer running an electronic health record. They are running Epic. Companies that have built their engineering rhythm inside Cursor or Claude Code have stopped using an AI tool and become Cursor companies or Anthropic companies. The cost of leaving is no longer technical. It becomes a question of who you have become.

Owning your execution layer is how you keep that question yours. When the platform your people work inside belongs to you, the model is a supplier, and your identity is your own. When it belongs to the vendor, the model is the least of what you have surrendered.

I do not write this from the outside. I pay Anthropic every month, I have shaped my own work inside their product, and I cannot imagine going back. The honest report is that the trap exists, the trap works on me, and the only protection is to know it is there and to let that knowledge shape what you build.

The executives who treat free as a procurement event in 2026 will stand in a different place three years from now than the ones who treat it as an IT experiment. They will have accepted most of the offers. The difference is that they kept the seam between the model and the work visible, they negotiated the exit on day one, and they owned the layer that compounds.

Rent the intelligence. Own the execution. Free is the most expensive word in enterprise technology, and the price tag is just on the next page of the contract.

References

1. Microsoft 365 Blog, Advancing Microsoft 365 (December 4, 2025), and the Microsoft 365 Copilot pricing page: Copilot Chat free tier (January 2025) and base suite price increases effective July 1, 2026.
2. Salesforce, Salesforce Introduces New Flexible Agentforce Pricing press release (May 15, 2025), and the Agentforce pricing page; SaaStr analysis of Agentforce deal and paid counts.
3. Slack Help Center, Updates to feature availability and pricing for Slack plans (August 17, 2025).
4. Atlassian support documentation, Rovo billing and licensing (April 14, 2025).
5. ServiceNow Q3 and Q4 2025 earnings calls and disclosures on Now Assist annual contract value and Plus-tier deal volume.
6. Anthropic raises $30 billion in Series G funding at a $380 billion post-money valuation (February 12, 2026): Claude Code run-rate revenue.
7. AT&T v. Broadcom filings, New York State Supreme Court (September 2024), as reported by CIO Dive and The Register.
8. Model Context Protocol: an open standard for connecting models to tools and data, adopted across OpenAI (March 2025), Google, Microsoft, and AWS, and contributed to the Agentic AI Foundation under the Linux Foundation (December 2025).
9. Open forks following vendor license changes: OpenTofu (from HashiCorp Terraform, 2023) and Valkey (from Redis, 2024), both hosted by the Linux Foundation and backed by major cloud providers.
10. Kristina Shampanier, Nina Mazar, and Dan Ariely, Zero as a Special Price: The True Value of Free Products, Marketing Science 26(6), 742-757 (2007).
11. Michael I. Norton, Daniel Mochon, and Dan Ariely, The IKEA Effect: When Labor Leads to Love, Journal of Consumer Psychology 22(3), 453-460 (2012).