Cheaper Models Won't Make Your Agent Cheaper. Ask S3 Why.

Start with the evidence, because it is simple. Willison's observation is that S3 "used to drop in price reasonably often" and that "there hasn't been a price drop in a full decade" (Willison, HN comment). His comment makes no claim about why. We shouldn't invent one either. The shape alone teaches a lot: an early run of cuts, then a long plateau.

Here is our reading, and it is analysis rather than reporting. Price cuts work as a recruiting tool. They matter most while customers are still choosing where to put their data. Once the data is in place, heavy and tangled into everything else, a cut mainly gives up margin on customers who were never going to leave. The service stays a commodity in engineering terms. Commercially, it stops acting like one.

Now apply that to AI agents. An agent that has run for six months is more than a model call. It holds a memory of your preferences. It holds the skills and workflows you installed or built. It holds credentials for your email, calendar, and files. It holds a history of corrections that makes it behave the way you want. That pile of state is your agent's data gravity, the same force that kept S3 customers in place.

If the pattern carries over, the next few years of agent pricing may look like early S3: visible price cuts while platforms compete for new users. After that could come a long, quiet plateau once those users have built up enough state that leaving feels expensive. The model underneath can keep getting cheaper through all of it. Your bill does not have to follow.

If lock-in explains the plateau, the defense is to keep the cost of leaving low. You cannot set a vendor's prices. You can check whether you are stuck. Four questions make up the test, and none of them requires writing code.

First: can you export your agent's memory in a form another tool can read? A download button is not enough if the file only makes sense inside the product that made it. Second: do your skills and workflows move with you? Suppose you rebuilt your morning briefing, inbox triage, and expense routine somewhere else. Would that take ten minutes or a weekend? Third: can you change the underlying model without rebuilding anything? If a leaner model ships next quarter, can you point your setup at it, or do you have to wait for your platform to decide? Fourth: can you see what each task costs? A flat subscription makes it hard to tell whether model price cuts are reaching you.

People who track their OpenClaw cost per month, or their bill on a hosted service, should keep a second number next to it: roughly how many hours it would take to leave. That second number is the one that decides whether the first can ever fall.

Timing matters. The test is cheap now because most setups are young. Every month of saved memory and every custom workflow raises the cost of leaving, and it rises quietly. S3 customers did not choose lock-in on a particular day. It built up. Run the test while failing it still costs little to fix. If you are comparing OpenClaw alternatives, rank them on these four questions before you look at the price.

The supply side is moving the way commodity markets usually move. TheSequence's September 27 issue framed that week's headline model release as "Opus 5.5 Gets Leaner" and described "a cheaper coding agent" (TheSequence Radar 940). The same issue asks outright whether AI compute is "a new commodity, or is it?" That question mark deserves attention, and we come back to it below.

Willison's year-in-review keynote, given as the closing talk at the WeAreDevelopers World Congress North America in San Jose, adds the other half of the pattern. On the year's releases he wrote: "As is usually the case with new models, these were incremental improvements on the models that came before them" (Willison, 2026 in LLMs). Steady small gains with falling prices is the textbook route to commodity. When the next model is a bit better and a bit cheaper, the model becomes the part you can swap out.

That is good news for anyone paying per token directly. For most people using agents, it is less relevant. You don't buy tokens. You buy an agent that sits between you and the tokens, and that layer is what the S3 pattern covers. Cheap inputs raise the margin available to the middle layer. They don't dictate what the middle layer charges.

TheSequence's question mark fits here. A commodity for the company buying compute can still show up as a premium product for the person buying an agent. Both can hold at once. Which one users see depends on how easily they can leave.

Stratechery got to the structural argument first, and credit belongs there. Its September 28 essay, "Apps, Agents, and Aggregation," opens by recalling how Steve Jobs pitched the iPhone as three products in one: "a wide-screen iPod with touch controls, a revolutionary mobile phone, and a breakthrough Internet communications device" (Stratechery, Apps, Agents, and Aggregation). The implication is that an agent is a bundle as well, and that whoever owns the bundle owns the customer. Aggregation Theory, in short: platforms win by aggregating demand and commoditizing supply.

We are not trying to reargue that. We want to add something the frame leaves open. Aggregation Theory predicts who captures value. It is quieter on what the winner does with prices once it has won. The S3 record fills that in. A winning aggregator doesn't have to raise prices to get richer. It just has to stop cutting them while its own costs keep falling.

Two of our house frameworks explain the mechanism. The Harness Hypothesis holds that the value in AI sits in the harness connecting the model to the world, not in the model. Commoditize Your Complement explains why harness makers are glad to see models get cheaper and interchangeable: every model price cut widens the harness's margin, provided the harness is the part you can't swap out. Whether you run a self-hosted setup or a hosted runtime like Claude Managed Agents, the incentive is the same. The harness wants the model to be replaceable and itself to be irreplaceable.

This is the core ai agent business model question, and it has nothing to do with which lab ships the best model. It comes down to which layer holds your context. Aggregation tells you who ends up holding it. The price plateau tells you what that will cost.

The layer that holds your context is also changing shape. The same TheSequence issue that reported a cheaper coding agent covered Meta shipping "glasses that can act on what you see" (TheSequence Radar 940). That is a new kind of agent surface. The context it collects is no longer typed prompts or linked inboxes. It is what you look at.

Meanwhile, Stratechery spent the same week on an interview with Colossus editor-in-chief Jeremy Stern about profiling Mark Zuckerberg (Stratechery interview). It is a fair read that analysts are paying close attention to the person running the company betting on that surface.

Why does this matter for portability? Ambient context is the hardest kind to move. You can export a list of saved preferences. It is much less clear what "export" means for months of an agent learning your routines by watching them. Every surface that gathers context passively rather than through explicit input makes the portability test harder to pass. The S3 pattern would predict a steeper plateau there.

There is also a pricing point hidden in plain sight. Stratechery itself sells a bundle: Stratechery Plus includes the Update, the Interviews, and podcasts including Sharp Tech, Sharp China, and Dithering (Stratechery interview). For a reader, that bundle is plainly good value. The pattern still generalizes: bundles hide unit prices. When an agent subscription bundles the model, the harness, the device, and the integrations, you can no longer see whether the cheaper model is reaching your bill. That is the fourth portability question, and bundling makes it harder to answer.

The best case against this argument goes like this. S3 is one product from one dominant provider. The agent market is crowded, with open-source harnesses, hosted runtimes, device makers, and every major lab competing. In a market that crowded, someone always undercuts. On top of that, models don't only improve in small steps. Willison notes that "every now and then when a model improves" the change is bigger than incremental (Willison, 2026 in LLMs), and step changes like that reshuffle markets before any plateau can settle in.

That objection is strong, and part of it is right. Competition does drive prices down. But the discount goes to the customers who can move. A new entrant's lower price means nothing to a user whose memory, workflows, and credentials are stuck in the old platform. The plateau isn't charged to everyone. It is charged to the people who are stuck. A crowded market with high switching costs can still produce flat prices for its long-standing users while it runs promotions for new ones.

The step-change point helps our argument more than it hurts it. When a major new model arrives, the users who get it first are the ones whose setup lets them switch models without rebuilding. That is the third portability question.

Open source deserves a specific note. In our Molt Cycle framing, open agent projects eventually reach a commoditization phase, and that phase tends to benefit users most. This only happens if the formats are open along with the code. An open-source harness that keeps memory in an undocumented format fails the portability test just as a closed one would.

One more pressure could push prices up instead of down. TheSequence's issue also covered Washington and Beijing discussing AI risk (TheSequence Radar 940). Any compliance costs that come out of talks like these would land on the middle layer, and customers who can't leave would be the easiest ones to pass them to. That is speculation. It points the same way as everything else, though: portability is what turns a cheaper market into a cheaper bill.