We are witnessing the end of the ‘All You Can Eat’ era for AI. For a year, we’ve been the competitive eaters at the subscription buffet, running loops and agentic sub-tasks on a flat $20-a-month fee. Anthropic just pulled the tray away, and Ben Thompson thinks it’s an obvious move.
A lot of people are understandably upset about this change, but it seems like an obvious one for Anthropic to make: yes, there are competitive concerns, particularly now that OpenAI owns OpenClaw (an acquisition that I haven’t written about that does make a bit more sense), but the bigger issue is the exponential increase in tokens that are caused by agents, particularly agents that aren’t tuned to the model. Subscription pricing for a good with meaningful marginal costs was always questionable; it’s completely untenable if you remove human friction from usage and replace it with an agent that never sleeps and has no incentive to increase efficiency. More generally, the reason to mention these three stories together is to note that AI generally and agents specifically are going to break a lot of things, above and beyond the security concerns I wrote about last week. The Internet’s removal of distribution friction broke a whole host of industries and business models; AI’s removal of substantiation friction is set to do the same thing, and tech companies and services are likely to be the first set of victims.
Source: OpenAI Buys TBPN, Tech and the Token Tsunami – Stratechery by Ben Thompson
Ben notes the ‘exponential increase in tokens,’ but the true second-order effect is what those tokens actually represent. It’s not just that LLMs require more compute; what they produce, from the logs, the loops, and the agentic exhaust, requires even more. From that perspective, when Anthropic decides that one cannot use their LLMs for non Claude harnesses, it has major downstream productivity impacts.
I am a high-burn statistical outlier. I run constant agentic loops on a flat-fee subscription. I am one of the very users who is affected by the changes at Anthropic.
Let’s revisit what happened last week:
Anthropic cut off the use of flat Pro and Max subscription limits on third-party harnesses like OpenClaw. Those users now have to use discounted extra-usage bundles or API keys instead. Anthropic’s stated reasons were capacity, sustainability, and the claim that these tools create usage patterns the subscriptions were not built for. It also offered a one-time credit equal to the monthly plan cost and said refunds were available.
The rest of the article attempts to see this from different viewpoints in an attempt to arrive at a reasoned conclusion.
To understand why this feels like a blueprint for user resentment, we have to look at the math Anthropic is using to justify the cage.
- Token explosion: The flat rate plans released when these chatbots were non reasoning models. Since then, we have the first exponential increase with
reasoningmodels. The second came in the form of coding agents / generally agents. These are harnesses that keep an active and sub loops of agents that are constantly churning through tokens. From a product perspective, that’s more value per $ for the user, but it’s economically unsustainable for Anthropic. - Protect quality for “code” users: Assume for a moment that 3P harnesses are less efficient than
Claude Codein using the models. Then, they want to protect the product quality forClaude Codeusers. This makes an artificial tie of the model to the harness but one can see the argument be logical. - Interface protection: Anthropic realizes that the harness is the interface to the model. To protect the business you have to aggregate demand and that means owning the interface of engagement. Their entire business model is dependennt on it. To be fair to Anthropic, imho, they have the best taste when it comes to interfaces. So, this is somehow to ensure that they don’t subsidize someone else’s interface.
- Consumers won’t care: We are in the heights of renting everything with major Terms of Service restrictions around how we use them. Nobody owns their music, their movies, or digital services anymore. It’s all subscriptions and in a world where everything is rented and comes with restrictions, consumers will just accept this as the state of the world.
The logic holds up in the boardroom but falls apart at the terminal. if the math is sound, the optics are a masterclass in how to alienate the very power users who built their momentum.
- Bait and Switch: Most people will see “includes
Claude Code”, “more usage” and “up to 5x - 20x for Max” usage and does not read that as limited to “Anthropic owned surfaces.” This is a blueprint for consumer resentment. It’s easy to include this statement in the pricing table. However, it’s going to be very hard to justify and worse, it is going to become a matter of comparison. Anthropic is likely betting on the fact that even OpenAI and Google will follow their pattnern and limit the usage of models to their harness. There is a darker pattern that can be employed here - one where you slowly introduce more friction - noisier ways to make the 3P harness experience worse than the 1P harness. It’s been done before - Printers claiming “non-genuine” ink until you buy the 1P cartridge. - Anthropic believes they can provide can build a 1P experience that will force the friction of the new deal into the background. To their credit, the multi-platform agent is a great step forward. I also inherently think they are approaching the security model in an honest manner. However, consumers inherently understand that there’s value in choosing how to use items they purchase for fairly valid use-cases (I am not talking about abuse here). This is a bit like if a high end ramen place claims that you can only truly appreciate the umami by using their proprietary, ergonomically-correct ceramic chopsticks. Any other utensil is ‘sub-optimal utensil friction.’ My own experience with third party harnesses often suggest that they are more token efficient (ymmv).
Benefits of the move
An opportunity to understand value
At the end of the day, when it comes to game theory: Anthropic made this decision because they could afford to and it sounds like they may have made the right call.
This move creates a clean test. If users keep paying and migrate to Anthropic’s native surfaces or paid usage bundles, Anthropic learns that Claude itself and Claude Code have standalone pricing power. If users churn, downgrade, or shift to rivals and bring-your-own-key setups, Anthropic learns that a lot of the old value sat in open orchestration rather than in the subscription bundle. It also stops hiding heavy token burn inside a flat fee and makes unit economics easier to see. That is a plausible inference, not something Anthropic has publicly admitted. The facts that support the inference are the new metered path for third-party use and Anthropic’s existing separation between consumer bundles and usage-based enterprise pricing.
The IPO angle
This is obviously speculative: Both Anthropic and their main competitor lab - OpenAI - are scheduled to go public in 2026. Anthropic last announced a $30B round at a post valuation of $380B while OpenAI announced a $122B round with a post valuation of $850B. Anthropic has the incentive to try and front run OpenAI’s IPO. Front running the OpenAI IPO forces OpenAI to defend its (likely) messier consumer-token-economics against Anthropic’s now cleaner narrative. Anthropic already has the advantage with enterprises and hence their monetization is much more straightforward. The pay as you use is for tokens and as harnesses cause token explosion, they are able to present a healthy narrative around unit economics and long term valuation. By front running, they also put OpenAI’s narrative in question, which is far more complex given their consumer bent.
While Anthropic polishes its narrative for Wall Street, the individual developer is left holding a more expensive, more restrictive bag. When the bundle breaks, the only thing left on the table is the bill. For the power user, the choice is no longer between ‘Pro’ or ‘Max’; it’s between agency and convenience.
Consumer options
As someone attempting to analyze the situation I understand why Anthropic made this move. As a consumer, things are more grim.
- Rely (and pay) Anthropic for the subcription only if native
Claude Codeis the value. This isn’t my case. - The labs will like redraw what is “included” in the subscription. OpenAI is moving at it from a consumer angle of “more value to increase subscription price later.” Anthropic’s approaching it from the point of view of heavier use enterprise and setting up token economics to match the infra.
- If so inclined, understand the seams and the layers: Just like when setting up a home network, you could just use the service provided all-in-one router, access point. I choose to split them apart - I provide my own router, my own switches and access points. This allows me to swap out systems for an upgrade and leverage the service provider for what they do best - provide the tubes for the traffic. By decoupling the agentic harness from the model and the context, I can reclaim the ability to swap the ‘tubes’ without losing the data. I am not just a customer, I am the architect of my own stack.
- The only way to maintain agency is to pay for the layers, not the bundle. Vote with your wallet to providers who have an open permission model. I’ve learned from the world of internet and mobile that prioritizing convenience over everything else ends up in being locked to providers.
And if I cannot at a reasonable cost, then maybe Anthropic’s offering is not the right one for me.
Holding two thoughts at once
Anthropic’s showing strong management - it’s found a leak with its subscription pricing and is plugging it with improved Claude Code capabilities and aligning its tokenomics. The optimistic read here is that they are committed to operating a sustainable business and are being honest to align pricing with costs. Taken seriously, this also suggests that this is likely the eventual outcome for Codex and Gemini.
However, as a consumer, this only makes me more resolute in building for the exit. I want a stack where the provider is a utility, not a cage.