The evolution of public access to frontier AI

In December 2024, access to frontier AI became an order of magnitude more expensive ($20 —> $200 per month) when OpenAI launched the ChatGPT Pro tier. Since then, the $100-$200 price range has become table stakes for using top AI models without constantly hitting rate limits.

This price increase makes sense. In 2024, the main “portal” into the AI world was a chat window. The AI was paused if you were paused, and most conversations were short back-and-forth exchanges. Then, in late 2025 to early 2026, AI usage became a completely different story. Demand for tokens skyrocketed as it became the norm, at least for enthusiasts, to have long-running agents working autonomously and in parallel. It felt like a battle between users of OpenClaw, Claude Code, and Codex to see who could use more tokens. If you weren’t calling your AI from a phone and setting them on a week-long endeavor of trading stocks through a workflow powered by eight sub-agents, topped off by flashing your room’s LED lights green every time a winning trade was made, you apparently were getting left behind.

Of course, this wasn’t the only way people were using agents. It really was a new, much more powerful way of using AI. For $200, you could have pretty much unlimited access to an autonomous intelligence that writes code, runs code, pushes commits to a repository, manages your files, formats your documents, does research, and serves as an adviser. A lot of busy-work became automatable.

In 2024, $20 for ChatGPT Plus was a great deal for most people. In 2026, paying $100 for Claude Max even felt a little cheap, especially when it became evident that Anthropic were having trouble keeping up with the new, more intense level of usage at this price point. Whatever compute they had available was getting squeezed away, and they had to take measures to save their GPUs. For example, their models became jarringly lazy. More effort levels were added to Claude Code as if to get you mixed up in the lower tiers. An “adaptive thinking” button was added, as if they were saying: “Now, you can finally enable your models to think for longer!” But actually, they removed the option to force their models to think, and enabling “adaptive thinking” just meant that you were allowing the model itself to decide whether to think or not. If you had it off, the model wouldn’t think at all. It became increasingly frustrating to beg a model to do what you want, especially because the capability was there, but it felt like some mysterious force in the background was making it dumber.

OpenAI looked to be doing better, and there was a moment where people got fed up with Claude Code and really considered switching to Codex. Snide remarks were coming out of OpenAI, not-so-discreetly implying that Anthropic underestimated compute demand and couldn’t keep up. This was sweet revenge for all of Dario Amodei’s talk about other companies in the industry (obviously OpenAI) recklessly making investments. It seemed like Anthropic were the ones that actually underinvested.

Despite OpenAI portraying themselves as being much better off, the level of usage of either company’s models was not sustainable at the $100-$200 price points. We were overdue for another price hike — one of these companies just needed something big to justify it.

That big thing came from Anthropic in the form of Claude Mythos. It was a step-change — a clear winner over competitors’ models. And the details surrounding its release, including but not limited to its price, indicate the arrival of a new era for public access to AI.

Pricing

Even users on the $200 plan won’t be able to use Fable 5 (the safer, general release of Mythos) through their subscription at all after July 7th. This is the first time a publicly available SOTA model cannot be accessed through any tier of subscription. Access is completely limited to the API.

In pure API costs, Fable 5 comes in at double the price of Opus 4.8. But this doesn’t mean overall costs will increase 2x for users making the switch from Opus to Fable; it’s been estimated that heavy users of the $200 Claude Max subscription would spend at least $1000 per month if all their usage was through the API, and likely much more. By making Fable API-only, Anthropic has raised the price for frontier AI by another order of magnitude.

$20 per month was something that most people in the U.S. could afford. $200 per month was something that most people in the middle-class or above could afford, and even at that price it’s perhaps the highest-value subscription you could have right now. But thousands of dollars per month, even for the ridiculous value that AI provides, is a whole new class of pricing. It does make sense: a new class of pricing for a new class of model. But it’s beyond what most individuals could pay, let alone would pay.

Right now, most people don’t really care. One reason is that most people work for a company, and it’ll be up to the company to bear the costs of paying for Fable so that its employees can use it. Most people aren’t trying to use AI to work for themselves.

Another reason is that it’s still not apparent to many people that frontier agentic AI is tangibly useful. It takes a lot of time, experimentation, and customization to get terminal agents really working for you. Even then, there’s still a real limit to their usefulness beyond coding tasks. As a result, most people either: a. don’t realize what can be done with terminal agents, b. know what terminal agents can do but don’t see enough value, or c. just don’t care about all this AI stuff. So most people aren’t moping around over the fact that they can’t afford to hammer the Claude Fable API.

But what happens if it does become obvious that people can benefit greatly from using a frontier AI model over the next best option, even if they’re not enthusiasts? If the #1 model is an order of magnitude more expensive than the #2 model, the prayers of our friends on X would be answered: those stuck with the #2 model actually would get left behind, not because they’re ignorant or unproductive, but because they just can’t afford the frontier.

Gate-keeping

For 2 months, Anthropic didn’t even release Mythos to the general public. Through Project Glasswing, they narrowed down the pool of users to a select few companies in order to “secure critical software for the AI era”. If it’s true that Mythos is highly capable (and willing) to find vulnerabilities in critical software, then it’s good that it wasn’t instantly released to everyone. But this choice is important because it marks the first time that a lab made access to a highly-publicized leading model exclusive. When Anthropic finally did make Mythos generally available two months later, it was through a safer version called Fable, which outright refuses to do most biology-related work and some coding-related work.

This makes sense from a safety perspective. It also makes sense from a business perspective. Why wait until you’ve made the model safe to launch for everyone, when you could make a bunch of money offering the initial version to wealthy organizations? Yet, this sets a precedent for a habit that can quickly devolve into reserving access to frontier models for the biggest customers, even for perfectly safe use-cases. Project Glasswing is already a form of this: I found no evidence that Anthropic’s partners only had access to Mythos’s cybersecurity capabilities, and later those companies got access to the innocent safe stuff at the same time as everyone else. Presumably, they had two months to use a Mythos-class model for all kinds of tasks before the general public. So, the Mythos release serves as evidence that a lab with a leading model can and will give certain groups a head-start with that model.

The government

The whole fiasco with the government forcing Anthropic to take down Fable must’ve hurt Anthropic. All that hype, all that opportunity to snatch users away from their competitors, just to get slapped in the face over a seemingly minor and indistinct incident. To make it worse, OpenAI released their answer to Fable 5 (GPT-5.6 Sol) just in time for when the government lifted the ban on Fable. Based on first impressions, Fable is still better. But the window where Anthropic has an undisputed king, such that even competitors’ customers are eager to switch, has closed. It was wasted.

It’s unclear exactly why the government banned Fable 5, but it was likely a mix of:

  1. Anthropic’s marketing (or as some put it, fearmongering) about Mythos’s ability to find vulnerabilities in critical software made the government more antsy than they should’ve been about foreigners using Fable.
  2. The government wanted to show Anthropic who’s boss after their previous conflicts.

This is another clear example of how public access to frontier models can be limited, and this time it comes from government intervention. As models get more powerful, the government will get more involved. In the pursuit of national security and preserving America’s lead, democratization of AI becomes less of a priority.

Where this leads

Overall, most of the progress in AI over the last year has been driven by data and agentic harnesses (e.g., Claude Code) rather than through differentiating architectural breakthroughs. OpenAI president Greg Brockman himself has stated that the reason Codex was initially behind Claude Code was that OpenAI trained models more on competitive programming data than messy codebase data. Chinese labs have even focused on distilling Anthropic’s models (training on Claude’s outputs) and have continued to closely trail the frontier. This points to data as having been the primary driver of recent progress in LLMs.

Owing to this, OpenAI and Anthropic have remained neck-and-neck for a while now, but this trend may not hold as proprietary architectural breakthroughs cause greater divergence. Whoever takes the lead as a result will have more freedom to price their models expensively. Furthermore, safety concerns and government pressure will push them to make access exclusive. This stratification of access to AI will create greater variance in what different groups can achieve.