There have been a lot of anxious murmurings lately about the price of AI going way up. Developers are spending more money training and running AI systems as competition thickens and the electrical grid gets squeezed. Customers, meanwhile, are shelling out more in order to get access to the latest models.
Earlier this week, Anthropic—moving towards what’s expected to be a historic IPO—released Fable 5, a dialed-down version of the secretive and supposedly extremely powerful Mythos. Fable 5 costs twice as much as its predecessor, Opus 4.8, even though some users have complained that the former’s touchy safety guardrails render it effectively unusable in some contexts. Presumably with its ear to those anxieties, OpenAI is now weighing significant reductions to the price they charge for tokens (the basic unit for measuring AI usage), the Wall Street Journal reported on Thursday.
To anyone who isn’t deeply immersed in the finer intricacies of AI finance, this can all be a bit baffling. It would be extremely convenient if there were some simple method for converting one million “input tokens” to a particular task, for example, but unfortunately, that’s not the case. Each task brings its own computational demands, which for pay-as-you-go models means users will have to pay different amounts depending on how they use AI. Subscription tiers offer a bit more simplicity, but these plans come with their own terms and prices, which vary between companies and models.
To clarify things, here’s what you need to know about the pricing models for three of the AI industry’s most powerful models:
Fable 5
First up, Anthropic’s latest release, the fabled Fable 5.
Subscribers to Claude Max, Pro, Team, and seat-based enterprise plans can use Fable 5 with their plan’s existing token allowances until June 23. Beginning on that date, the company plans to revert to a pay-as-you-go model for all Fable 5 users, meaning the more intensively they use the model, the more customers will have to pay—regardless of whatever subscription tier they might be subscribed to.
Anthropic intends to reinstate the usual subscription-based token allowances for Fable “when sufficient capacity allows us to do so,” according to a blog post published earlier this week. It’s not yet clear what will happen to paid Claude subscribers who haven’t exhausted their entire token allowance before the June 23 cut-off date; we’ve reached out to the company for answers and will update this story as soon as we know more.
The key thing to remember here is that Fable 5 consumes more tokens than Anthropic’s earlier models. So if you’re currently paying $100/month for the Max 5x plan, you’ll continue paying that same amount using Fable, but there’s a good chance you’ll hit your token limit faster.
Starting on the 23rd, all users will need to pay $10 per million input tokens and $50 per million output tokens when using Fable.
According to one common arithmetical shortcut, one token translates to roughly four characters of text; it therefore takes a hell of a lot of written prompts to get to a million tokens, meaning you can get plenty of value out of $10 if you’re only using Fable to, say, write work emails or generate dinner recipes. Then again, if that’s all you need AI for, you may as well use a free chatbot: Using Fable to respond to simple text chats is like driving a McLaren W1 to drive to your next-door neighbor’s house.
Fable 5 specializes in long-running autonomous tasks, like writing software code, which requires many, many more tokens—we’re talking in the hundreds of thousands to millions for both inputs and outputs. Your monthly bill will therefore be significantly higher than it would be if you were just feeding simple text prompts to the model. But if you’re already paying, say, $200/month for the Max 20x plan, you may not be paying all that much more for usage credits than you are already: Using 10 million input tokens and 5 million output tokens would lead to a bill of $350 (($10 x 10) + ($50 x 5)).
The price of using Fable 5, in other words, will depend entirely on the demands of the tasks you’re using the model for—that’s, of course, the basis of the pay-as-you-go model. If you tend to hand models complex tasks that require many steps and long stretches of time, proceed with caution.
GPT-5.5 Pro
Released in April, GPT-5.5 Pro is the latest model powering ChatGPT. It’s available via OpenAI’s Pro plan (which costs $200/month), and also on the company’s Business ($30/user/month) and Enterprise (custom pricing) tiers.
Developers using GPT-5.5 through OpenAI’s API, on the other hand, are charged by a pay-as-you-go model like the one that will start applying to Fable later this month. At $5 per million input tokens and $30 per million output tokens, it’s significantly cheaper than Fable (and just a little bit more expensive than Anthropic’s second-most valuable publicly available model, Opus 4.8) It also comes with 50% cheaper batch tokenization option, which essentially allows OpenAI’s servers to handle bundles of similar requests in single “batches,” boosting computational efficiency but also resulting in slower response times.
Gemini 3.5 Flash
Google has highlighted what it has said is a unique blend of speed and agentic capabilities with the most powerful version of Gemini, 3.5 Flash, which dropped last month.
It’s available for free with usage limits, and developers can build off the API for $1.50 per million input tokens and $9 per million output tokens—by far the most affordable option of the three models we’ve looked at thus far.
The bottom line
Just as there’s no standardized, industry-wide pricing model for AI, there’s also a huge amount of variation in the benefits and drawbacks of each model.
For many users who just need a chatbot to serve as a glorified search engine, the free versions of Claude, ChatGPT, or Gemini are probably fine. Anyone whose work demands a more advanced model for, say, coding or research purposes is probably better suited to pay for a subscription. Just pay attention to the fine print before you make a choice, and be on the lookout for key phrases like “usage limits” and “pay as you go.”
Read the full article here
