Price hike for AI coding tools: The Free Lunch Is Over

Contents

Services such as Cursor, GitHub Copilot, Anthropic Claude Code, or Google's Gemini CLI help developers to write and debug code faster and automate tasks. However, a clear trend has emerged in recent months: providers are tightening the price screw. Generous flat rates are increasingly giving way to usage-based models and stricter limits. In some cases, power users are incurring considerable additional costs.

Rainer Stropek

Rainer Stropek ist IT-Unternehmer, Softwareentwickler, Trainer, Autor und Vortragender im Microsoft-Umfeld. Er ist seit 2010 MVP für Microsoft Azure und entwickelt mit seinem Team die Software Time Cockpit.

This article examines the current developments in the pricing models of the tools mentioned, the reactions of the community, and the question of what this means for developers and companies.

Cursor: From flat rate to credit system

The editor Cursor from the start-up of the same name was long regarded as one of the best AI assistant editors with a fair, predictable pricing model. Until recently, the Pro subscription for 20 US dollars per month included a quota of 500 AI requests, with additional requests available for an extra charge. However, Cursor changed this simple model rather abruptly a few months ago, leaving many people scratching their heads.

Without much notice, Cursor abolished the fixed request limit and instead promised “unlimited use” with dynamic rate limits. What sounded like good news quickly turned out to be a coverup for a massive cutback. What used to include 500 requests suddenly became only around 225 at the same price.

Moreover, the unlimited use only applied if you allowed the system to automatically select a model. Anyone who specifically selected a particular model, such as Claude Sonnet 4, had to pay for each request from a limited credit balance. The community saw this procedure as a breach of trust and openly spoke of misleading. In Reddit discussions, there were accusations that Cursor had promised “unlimited” but tacitly limited it.

Many also criticized the fact that Cursor delivered significantly worse results under the new rate limits. Among other things, one post stated that the AI assistant often stopped in the middle of a task in limited mode or only completed partial tasks, requiring several follow-up questions. Users have expressed the suspicion that Cursor deliberately rations the output so that the included credit is not used up too quickly. However, if you switch to genuine usage-based billing, the model suddenly works as usual again, but at a correspondingly higher cost. This has caused a lot of resentment among paying customers.

Power users of Cursor were particularly frustrated. Those who used the tool intensively reached its limits much sooner. In addition to the usage limits, the agent function was also a reason for this. This mode breaks down complex tasks into sub-steps but also means that a single command triggers dozens of API requests in the background and consumes masses of tokens. “Cursor has changed the formula with its agent so that huge amounts of credits are consumed,” reports a user on Reddit. This change led to some users using up their monthly quota in just a few days.

In response to the bottlenecks, Cursor offered new, more expensive subscription levels for frequent users but hardly communicated this transparently. Suddenly, a Pro+ plan for 60 US dollars (three times the limit) and an Ultra plan for 200 US dollars (twenty times the limit) appeared. The fact that existing customers had to pay more to get back to what was originally offered understandably annoyed many. The term “enshittification” for platforms that worsen their services over time to increase profits is an obvious one.

Cursor has since admitted that the price change did not go well. On X, the founders explained that they had listened to the feedback and missed the target. The fact is that the trust of many former fans has been shaken. Numerous users have migrated to alternatives or are considering doing so.

GitHub Copilot: Premium limits and pay-as-you-use

For a long time, GitHub Copilot was considered an “all you can eat” service for ten US dollars a month. The subscription included unlimited code completions and chat responses. Many initially wondered how Microsoft intended to make the service economically viable. And indeed, the turning point came in June 2025: GitHub introduced a limit for Premium Requests, which is roughly equivalent in price to what was previously considered unlimited. Since June 18, 2025, every paying Copilot customer has had a quota of premium requests per month, the scope of which depends on the specific price plan. Anything above this costs extra.

GitHub defines Premium Requests as all requests for advanced models and features outside of OpenAI's GPT model series. In particular, this includes the use of the Claude Sonnet model within Copilot, which is particularly popular for coding. “Do you want to use Claude or Agent? Counts as a premium”, one user sarcastically summarized the procedure.

Normal coding with the basic models (GPT-4.1 and GPT-4o) remains unlimited, but all more sophisticated AI functions count towards the limit of the respective price plan. If the limit is exceeded, 0.04 US dollars will be charged for each additional request, or Copilot will discontinue the expensive services if no additional budget is released. Microsoft has integrated budget management into GitHub specifically for this purpose: Team admins can set a monthly cost limit and optionally specify that Copilot stops automatically once the budget is reached. By default, the additional budget is set to 0 US dollars, meaning that you cannot spend more than the price of the basic subscription without actively opting in.

For power users, there is also a new Copilot Pro+ plan for 39 US dollars per month, which includes 300 instead of 1,500 Premium Requests. Business and Enterprise customers also have higher quotas according to their license but must also pay for additional usage.

Many existing customers feel that they have been left out in the cold. Microsoft's communication in particular has been criticized. For a long time, Copilot Pro was advertised as “unlimited”. The sudden introduction of a 300 limit is therefore perceived by many paying users as a bait-and-switch offer. The sarcastic Reddit post on the premium model quoted above received a lot of encouragement. Users also criticized the fact that GitHub offered hardly any transparency when rolling out the limits. Initially, there was no reliable way to view your usage. The fact that many only received the message “No usage found” when they tried to call up their statistics increased mistrust. In the official GitHub forums, comments such as “It feels like Pro has just been cut to make Pro+ more attractive” piled up.

For professional users, the change primarily means more planning effort and possibly higher costs. Although GitHub advertises that all paid plans continue to include unlimited completions and chat, anyone who wants to exploit the full potential of Copilot, for example with the Claude models in Agent Mode, will not get far with 300 requests per month, i.e., just under 15 requests per working day. Many developers now regularly reach this limit.

With the budget feature, Microsoft at least makes it possible to consciously control costs. Companies can monitor Copilot usage in this way. Ultimately, however, there is no way around digging deeper into your pockets for more intensive use, either by switching to more expensive plans or using the pay-as-you-go model.

The mood in the Copilot developer community is divided. Some realize that the AI flat rate was not economically viable. Others, however, are frustrated by the deterioration of the offer. It is noticeable that Microsoft communicated the introduction of the limits more actively than Cursor or Anthropic. There were blog posts, pages in the documentation, and information in the FAQs. Nevertheless, it remained unclear to many what exactly counts as a premium. For example, it is not clear from the documentation that Copilot's agent mode often executes many individual requests internally, so-called tool calls, and how these calls affect the quota of premium requests. For power users, the changes to the GitHub Copilot pricing plan mark the end of carefree Copilot use.