$10. For two prompts.
That's what one developer on Cursor's Enterprise tier spent in a single session last week - one prompt on GPT-5.5, one on Claude Opus 4.6-thinking. The month before, the same person burned $80 in a single week using Claude Opus 4.7, even after a 50% launch discount. This isn't a billing glitch or unusual usage pattern. It's where frontier AI pricing is heading, and the implications for tools like Cursor are straightforward: open-source models are about to get a lot more serious attention.
The Subsidies Are Gone
For the first few years of the AI tool boom, companies like Anthropic and OpenAI priced access to their best models below what it actually cost to run them. They were building habits, not margins. That period is over.
Frontier models now cost what they cost. Claude Opus 4.6-thinking runs "extended thinking" - meaning the model reasons step-by-step through a problem before answering, which is computationally expensive and produces better results on hard problems. Claude Opus 4.7 pushed capability further. Both are genuinely powerful for complex coding tasks. They're also genuinely expensive to serve, and when Anthropic stops discounting, that cost lands on whoever is paying the API bill.
For tools built on top of these APIs, this creates real pressure. Enterprise plans assume predictable usage. When two prompts cost $10, the math only works if most users stay well under that rate. Professional developers, who are exactly the people buying Enterprise tiers, don't.
Where Open Source Actually Holds Up
The honest answer is that open-source coding models have gotten good enough for most of the work. Models like Qwen, DeepSeek-V3, and the latest Llama variants handle autocomplete, refactors, boilerplate generation, and standard debugging at a fraction of the API cost of Anthropic or OpenAI's frontier offerings - often 5x to 10x cheaper depending on the provider.
Frontier models still pull ahead on genuinely hard problems: multi-file architectural changes, subtle concurrency bugs, novel design decisions where reasoning quality matters a lot. For that work, paying frontier rates is defensible. The question developers are now asking is what percentage of their daily coding falls into that category. For most, it's well under half.
Cursor already supports custom model configurations and local model connections. The technical option to route different tasks to different models has existed for a while. What's changing is motivation. At $5 per session on a frontier model versus $0.50 on a well-tuned open-source alternative, developers pay attention to which tasks actually need the expensive one.
What This Means for the Market
The tools caught in the worst position are those whose entire value proposition rests on giving users access to GPT-5 or Claude Opus at a flat price. As model costs rise and subsidies end, that business model compresses. Either pricing goes up and users push back, or margins shrink until they don't make sense.
Open-source models won't replace frontier options for every use case. For the hardest 20% of coding problems, the performance gap is still real. But for routine daily work, the gap between a top open-source model and a frontier one is now small enough that the 5-10x price difference makes the choice obvious. The developers doing the math right now are going to move first - and there are a lot of them paying attention.