If you want to feel the dizzying speed of AI's impact on our industry, here is a crazy number: Anthropic launched its Max accounts only a year ago. Claude Code, as we know it, only just blew out one candle on its birthday cake. Back in early 2025, autocomplete-on-steroids like Cursor was still the norm.
Claude and Codex were not the first coding agents, but came with an impressive innovation: the all-you-can-eat buffet. The Max plan gave almost unlimited use for a fixed price. It's crazy to look back at this, but the first mention of the $100 or $200/month price tag was met with laughter. Who on earth would pay that much?
Today, it's generally considered a steal. The productivity gains of a single subscription strongly outweigh the price tag. Having broad access to AI without having to worry about tokens spent is an extreme value for money and most likely, too good to be true.
Anthropic has been cracking down on "unfair use" by banning paying customers who systematically shovel too much steak on their plate at the buffet. OpenAI's latest model burns through its 5-hour limits faster than ever before. GitHub Copilot will switch from all-you-can-eat to pay-as-you-go in a few weeks. Claude Design gives you around 45 minutes per week of usage.
There are 3 trends converging that lead to the end of this business model.
The war for compute
AI uses massive amounts of GPUs, and those investments are a tricky balance. On the one hand, you don't want to fall short on compute. That will stunt your growth. On the other hand, investors get queasy when they see the huge CAPEX on the slide deck. Timing is also crucial. Do you order a new data centre when RAM prices are at an all-time high? Do you risk investing in a chip that might be outdated tomorrow?
Companies like Anthropic are victims of their own success, and their planned data centre growth is currently unable to keep up with users' demand. In 2026, they need to get selective with their compute.
Models have exploded
That first Claude Code back in 2025 came with Sonnet 3.7; Opus 4 wasn't out yet. Codex, similarly, launched on codex-1, a coding-tuned version of o3. Both were fantastic for their time, but they can't keep up with the frontier tech we use today. Over the last year, we've witnessed a new category of models doing things we couldn't imagine. GPT-5.5 and Opus 4.7 are amazing, but they are also much more compute-heavy than their predecessors.
When people complain that their Max plan has become slower and more limited, they forget to mention that the all-you-can-eat buffet has switched from serving pizza to Wagyu steak.
They've hooked enough of us
Software engineers are notoriously cheap. It's remarkably hard to get software developers, people paid to write software, to pay for software. It's part of the reason why open source is so popular.
So, the low entry of a GPT-Plus plan is ideal for this scenario. It allows engineers to get used to these tools and experiment without worrying about a price tag.
Today, most engineers are hooked on their coding agent. From a business perspective, that makes it the ideal moment to revert to pay-as-you-go.
The end of the all-you-can-eat buffet might seem like bad news for engineers, but it isn't. It opens up the market for a lot of innovation. We use Codex because the tokens are free, not because it's the best CLI agent. We use Opus 4.7 to generate meeting minutes because it's included in our Max plan. That's the equivalent of driving a Maserati to the supermarket.
When developers need to be mindful of their token spending, we will see a lot of innovative changes. We will see asynchronous lights-out engineering pipelines that optimise for GPU load rather than the real-time conversational UIs of today. We will see more local models. We'll use different models for different purposes. Opus can come up with the marketing plans, but Haiku will generate the e-mails. The competition between models and coding agents will get fierce as their consumers can now decide to switch to a better or cheaper competitor.
We'll move from all-you-can-eat to à la carte, and the industry will be better for it.
