# Why do AI prices keep falling while top plans cost more?

Date: 2025-06-25 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/ai-prices-keep-falling-while-top-plans/

**Answer:** The price of a given level of AI keeps dropping because newer, smaller and more efficient models do the same work for a fraction of the cost. Top plans cost more because they sell something different: the newest models, reasoning modes that use far more computing per answer, and much higher usage limits for heavy users.

Two price stories run side by side in AI this year. Developers who pay per use keep seeing cheaper models arrive. Meanwhile the most expensive consumer subscriptions have climbed to $200 a month and beyond: OpenAI's ChatGPT Pro at $200, Anthropic's Claude Max at $100 or $200, and, since Google I/O in May, Google AI Ultra at $249.99. Both stories are true, and they don't contradict each other once you see what is being priced.

## Two ways AI is sold

**Per token, through an API.** An API (application programming interface) is how programs call a model directly. You pay for **tokens**, the pieces of words a model reads and writes; OpenAI counts roughly 750 English words per 1,000 tokens. Prices are quoted per million tokens, with output usually costing more than input.

**Per month, through a subscription.** You pay a flat fee for a chatbot app, and the plan sets which models you can use and how much. Here the price buys access and limits, not a measured amount of text.

## Why per-token prices fall

The same level of ability keeps getting cheaper to deliver. Each new generation of models tends to include smaller, more efficient versions that handle everyday tasks well. OpenAI's own price list from July 2024 shows the gap:

| Model on OpenAI's July 2024 price list | Input, per million tokens | Output, per million tokens |
|---|---|---|
| gpt-4-turbo (older flagship) | $10.00 | $30.00 |
| gpt-4o (newer flagship) | $5.00 | $15.00 |
| gpt-4o-mini (small, efficient model) | $0.15 | $0.60 |

For many tasks, such as summarizing, classifying or drafting short replies, the small model is good enough, at a few percent of the older flagship's price. The same page lists a Batch API, which returns results within 24 hours at half price, for work that doesn't need an instant answer.

## Why top plans cost more

At the top end, three things push in the other direction.

**Reasoning models use more computing per answer.** A reasoning model works through a problem in steps before replying. As TechCrunch noted when Anthropic released Claude 3.7 Sonnet, its first such model, this uses more computing power in exchange for more reliable answers on some questions. A plan that offers lots of that is expensive to run.

**Heavy users want far more capacity.** Claude Max is sold directly as usage: the $100 tier gives five times the rate limits of the $20 Claude Pro plan, and the $200 tier gives twenty times. ChatGPT Pro, at $200, offers unlimited usage under OpenAI's terms.

**Bundles of new tools.** Google AI Ultra combines the Veo 3 video generator, the Flow video app, early access to Gemini 2.5 Pro's Deep Think mode, higher NotebookLM limits, agent tools, YouTube Premium and 30 TB of storage. Part of the price pays for things that are not chat at all.

So the falling prices describe a fixed level of capability, and the rising prices describe the frontier plus volume. Most people need much less of the frontier than the top plans offer.

## Which plan do you actually need?

Start from what you do, not from the plan names. Here is an illustrative way to match common needs to a tier; check each service's current page for exact limits, which change often.

| How you use AI | Likely fit |
|---|---|
| A few questions a week, short drafts, trying things out | A free plan |
| Daily writing, research or study, with file uploads and the main models | A standard paid plan, around $20 a month (Claude Pro is $20) |
| You hit the standard plan's limits most days, or rely on long reasoning sessions for work | A higher tier such as Claude Max at $100, or one of the $200 plans |
| You need video generation, large storage or a bundle of Google tools | Google AI Ultra, if those extras are worth it to you |
| You are building an app or automation | The API, paying per token |

The API row is often overlooked. Take an illustrative example: a weekly job that summarizes 200 customer emails, each about 2,000 tokens long, into 300-token summaries. That is 400,000 input tokens and 60,000 output tokens. At the gpt-4o-mini prices above, it comes to about 6 cents for input and under 4 cents for output, roughly 10 cents a run. Paid per token, a small regular task can cost far less than any subscription.

> **Try it:** For one week, note each time you use an AI assistant and whether you hit a limit or wished for a better model. At the end of the week, use the table above to pick the lowest tier that would have covered every entry.

Two habits keep the bill honest. Upgrade only when you actually hit limits, not in anticipation, and set a reminder before each renewal to ask whether last month's use justified the price. If you pay through an API, set the monthly budget and email alert that OpenAI's account settings provide, so a mistake in your code can't run up a large bill.

## Sources

- Pricing (OpenAI API, July 2024): https://openai.com/api/pricing/
- Anthropic rolls out a $200-per-month Claude subscription: https://techcrunch.com/2025/04/09/anthropic-rolls-out-a-200-per-month-claude-subscription/
- Google AI Ultra: You'll have to pay $249.99 per month for Google's best AI: https://techcrunch.com/2025/05/20/google-ai-ultra-youll-have-to-pay-249-99-per-month-for-googles-best-ai/