Anthropic cuts third-party access and changes the rules for autonomous agents on Claude
Anthropic made a move over the weekend that caught a lot of people off guard. The company blocked the use of Claude by third-party tools under the monthly subscription plan, and the ones who felt the impact immediately were users of OpenClaw and other solutions running autonomous agents continuously. The announcement came on Friday from Boris Cherny at Anthropic, posted directly on X, and it didn’t take long for the community to react with a lot of intensity.
The decision raises a question that goes beyond the technical side: how far can the current business model sustain the level of usage that agentic AI demands? Because this isn’t just a billing change — it’s a clear signal that the rules of the game are shifting, and power users are the first to feel it. 🤖
Why this change matters so much
This move by Anthropic exposes a growing tension at the heart of the artificial intelligence boom. On one side, power users want autonomous agents that run around the clock, executing complex tasks without constant human oversight. On the other, AI labs are trying to control costs, compute capacity, and how their models are being used. This friction between freedom of use and financial sustainability isn’t new, but it just became way more obvious.
As AI product manager Aakash Gupta put it in a post on X: the 20-dollar-a-month all-you-can-eat buffet just closed. The line is provocative, but it captures exactly what happened. Until now, the Claude monthly subscription worked like an all-access pass for anyone who knew how to exploit the system, and third-party tools were doing exactly that at a scale Anthropic could no longer afford to support.
What actually changed with Claude
Before this decision, many developers and power users were leveraging the Claude subscription plan to fuel external tools, especially those running autonomous agents continuously and automatically. This type of usage is very different from simply opening a chat and asking a question. An autonomous agent can fire off hundreds of calls to the model in sequence, handling complex tasks like research, data analysis, workflow automation and much more, all without direct human intervention. The resource consumption in this scenario is on a completely different level compared to typical conversational use.
The core problem Anthropic identified is exactly this imbalance. A regular user chatting with Claude a few hours a day consumes a fraction of the resources that an autonomous agent running in a loop consumes over the same period. When third-party tools like OpenClaw started using the monthly subscription as a foundation to power these agents, the pricing model simply couldn’t hold up anymore. The math doesn’t work for Anthropic, and the solution was to clearly separate these two types of usage.
In practice, here’s how the change works:
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
- Users can still access Claude models, including Opus, Sonnet, and Haiku, through external agent frameworks.
- However, they now need to pay through the Anthropic API or through a new pay-as-you-go billing system called extra usage, rather than relying on flat-rate subscriptions.
- This significantly changes the financial picture for developers and companies that built solutions on top of the subscription plan, since cost per token can scale quickly depending on the volume of operations the agent runs throughout the day.
What is OpenClaw and why it’s at the center of this story
For those unfamiliar, OpenClaw is a pretty popular open-source tool that lets users run autonomous AI agents continuously. Unlike a one-off chat conversation, these agents can operate 24 hours a day, seven days a week, executing long chains of reasoning and automated actions. The volume of tokens this kind of usage burns through is incomparably larger than that of a casual user.
The 20-dollar monthly Claude subscription had, in practice, allowed these super users to access high-cost models at a fixed and low price. It was a loophole that worked great for developers but created a serious sustainability problem for Anthropic. By pushing these developers toward API billing and paid add-ons, the company gains much tighter control over pricing, usage limits, and its profit margins.
The API as the official path for agentic use
Anthropic‘s move makes it clear that the API is the official and expected channel for anyone who wants to build on Claude in a serious and scalable way. This makes total sense from both an engineering and a business standpoint. The API offers granular control over every call, detailed consumption metrics, configurable limits, and a billing structure that directly reflects how much each application consumes. For Anthropic, this means revenue predictability and the ability to scale infrastructure based on each client’s actual demand.
For developers, the transition to the API comes with additional responsibilities. You need to manage access keys, monitor token consumption, implement retry logic, and deal with rate limits. Anyone already working with LLM APIs knows this drill well, but for those who were using the subscription plan as a shortcut to avoid all that complexity, the change represents a learning curve and a budget adjustment that can be significant depending on the project.
But there’s an upside here. Using the API opens the door to features the subscription plan simply doesn’t offer, like deeper integration with external systems, model behavior customization through more elaborate system prompts, access to specific versions of Claude, and the ability to use features like tool use in a much more controlled and auditable way. For anyone building real products, the API is the natural path forward.
Compute capacity is finite and super users were overloading the system
A fundamental point that often gets overlooked in this discussion is that compute capacity is a finite resource. Anthropic had already been implementing measures to deal with this even before blocking third-party tools. Stricter usage limits were already in place, including five-hour session caps during peak hours. These measures exist because heavy autonomous agent usage can reduce service availability for everyday users who just want to use Claude for regular tasks.
As Cherny himself explained: capacity is a resource we manage carefully, and we are prioritizing customers who use our products and our API. The message is straightforward. Anthropic wants to protect the experience of its paying customers within the official ecosystem, not indirectly subsidize massive usage through third-party tools exploiting loopholes in the pricing model.
The agentic AI business model is being stress-tested in real time
This decision by Anthropic is an important barometer for the entire industry. The rise of autonomous agents is putting enormous pressure on pricing models that were designed for much simpler and more predictable usage patterns. AI companies have invested billions in training and infrastructure, and the challenge now is figuring out how to monetize sustainably a type of usage that’s growing exponentially and has completely different characteristics from traditional chat.
The flat monthly subscription business model works really well when user behavior is relatively uniform. You pay a fixed amount, use the service within an expected range of consumption, and everyone’s happy. But when autonomous agents enter the picture, that uniformity goes right out the window. A single well-built agent can consume in one day the equivalent of what hundreds of regular users would consume in a month. Keeping that kind of usage inside a flat-rate plan is financially unsustainable for any company, regardless of size.
Agentic systems can run for hours straight and execute actions across multiple apps and platforms. This means AI providers end up footing the compute costs generated by super users who consume resources far above the average. As Cherny put it pretty clearly: Claude subscriptions were not built for the usage patterns of these third-party tools. 💡
The open-source community’s reaction and what comes next
The developer community using Claude reacted with a mix of frustration and understanding. Those caught in the middle of ongoing projects, especially ones relying on tools like OpenClaw, felt the impact in a very tangible way. Broken pipelines, unexpected costs, and the need to refactor integrations are real problems that don’t just disappear overnight. Anthropic’s communication via X helped provide some context, but the speed of the change left very little time to adapt.
The change is also fueling a significant backlash from the open-source community. Peter Steinberger, creator of OpenClaw who was hired by OpenAI, said he pushed Anthropic to reconsider the decision and shared alternative workarounds after the change went into effect. Some users are exploring the possibility of running models locally as a way to completely avoid usage limits imposed by cloud providers.
On the other hand, many more experienced developers acknowledge that using a monthly subscription to power high-volume autonomous agents was a loophole that was going to be closed sooner or later. The model was never designed for that, and exploiting the loophole worked until it didn’t. Now that Anthropic has formalized the rules, at least there’s clarity about what’s allowed and what isn’t, which in the long run is better for anyone building serious products on the platform.
Efficiency versus growth: the strategic divide in the AI industry
This move by Anthropic also reflects a broader strategic divide in the artificial intelligence industry: efficiency versus growth at all costs. Critics have argued that rivals like OpenAI, led by Sam Altman, lack financial discipline, while Anthropic has been presenting Wall Street with a more capital-efficient approach.
A former employee who worked at different cutting-edge AI labs told Axios that Anthropic emphasizes efficiency in how it trains and operates its models. The mindset at OpenAI, on the other hand, is reportedly that Altman could always raise more capital to sustain the scaling of compute capacity. These two philosophies represent very different paths for the future of the industry, and decisions like this one on Claude show in practice how Anthropic’s approach materializes.
What to expect going forward
What we’re seeing right now is the market adjusting in real time. Anthropic isn’t the only company that’s going to need to make decisions like this. As agentic usage becomes more common and sophisticated, all major players will need to rethink how they charge for their models. Per-token billing, while less predictable for the end user, is probably the fairest and most sustainable mechanism for this type of usage, because it directly reflects the actual compute cost of each operation.
Expect more friction like this in the near future. Agentic AI is significantly more expensive to run than traditional chatbots, and someone has to foot the bill. The speed at which agents are becoming more capable is making the business model, not the technology itself, the main bottleneck for the industry’s evolution.
The takeaway here is that building on top of consumption plans that weren’t designed for your use case is always a risk. The API with per-token billing might seem more expensive at first, but it offers something far more valuable for anyone developing products: predictable behavior, clear terms of use, and a support channel appropriate for the level of usage that autonomous agents demand. Anthropic’s message is straight to the point: if you want to build with Claude for real, the API is the right place to do it. 🚀
