Over the last two years, the economics of Generative AI have shifted. If you build with APIs, like we do in Theta Assist, the cost of calling the leading models has fallen by 80-95%. But if you buy user licences for products like ChatGPT Plus, Microsoft Copilot, or Google’s Gemini, you’re still paying the same monthly rate (or more) as when these services first launched.
This divergence is stark: the unit cost of AI is getting cheaper, but the subscription fees we pay as end-users are not. Let’s look at the detail and alternative approaches.
Note: we’ve used USD pricing in this as that is easiest to get.
OpenAI
- APIs: When GPT-4 launched in March 2023, it cost $0.03 per 1,000 input tokens and $0.06 per 1,000 output tokens. In 2024, OpenAI introduced GPT-4 Turbo and then GPT-4o, cutting prices by more than 80%. Today GPT-5 costs just $0.00125 per 1,000 input tokens and $0.01 per 1,000 output tokens – a 95% drop from the original GPT-4.
- Licences: ChatGPT Plus has been $20/month since February 2023. The model behind it has improved dramatically (now GPT-5), but the sticker price hasn’t budged.
Anthropic
- APIs: Claude v1 launched at a similar level to GPT-4. By 2025, Claude 2.1 sits at ~$0.011 per 1,000 input tokens and ~$0.033 per 1,000 output tokens. Lighter models like Claude Instant and Haiku bring costs right down to fractions of a cent, more than 90% cheaper than v1.
- Licences: Claude Pro launched at $20/month in 2023 and is still $20/month today. Heavier-use tiers (like Claude Max at ~$100/month) have been added, but the core plan remains fixed.
- APIs: In early 2023, Bard was free, with no API. By 2025, Gemini models are the cheapest on the market. The “Flash” variant is priced at ~$0.10 per million input tokens and ~$0.40 per million output tokens – that’s $0.0001/$0.0004 per 1,000 tokens. Virtually free at scale.
- Licences: Google went straight to enterprise pricing. Duet AI (now Gemini for Workspace) was $30/user/month at launch in 2023 and is still $30 today. A $20 Business tier was added in 2024, but the enterprise price point hasn’t shifted.
Microsoft
- APIs: Azure OpenAI has always mirrored OpenAI’s own pricing. As GPT-4o arrived and costs dropped, Azure followed suit. A GPT-4 call that cost $36 per million tokens in 2023 is now closer to $4 per million – an ~80% fall.
- Licences: Microsoft 365 Copilot launched at $30/user/month in mid-2023. It’s still $30. A consumer “Copilot Pro” tier at $20/month arrived in 2024, alongside GitHub Copilot at $10/month (unchanged since 2021).
The contrasting story
Here’s the story in summary numbers (all USD):
What are the alternatives
- Stick with continued high costs: If you’re buying licences, expect vendors to hold the line on per-seat pricing while justifying the same monthly fee with better models and more features.
- Reduce rollout: if you want to use these tools but reduce cost, only handout licences to the people you think will benefit the most
- Choose a different product: On the flip side a product based on API use – like Theta Assist - can be a much more economic option as the fees are banded, not per user and much lower than other comparable products.
In the end this comes down to fairness and access. Users have very different usage profiles and paying a high price for all users doesn’t make a whole lot of sense (unless you are the vendor). There are other options – contact us to learn more about Theta Assist.
Sources:
- OpenAI API Pricing Announcements (2023–2025) https://openai.com/api/pricing
- Anthropic Claude Pricing Documentation
- Google Workspace & Cloud AI Announcements
- Microsoft 365 Copilot Official Blog Posts
- Andrew Ng, The Batch – Falling LLM Token Prices (Aug 2024) (context on GPT-4 cost drop) - https://www.deeplearning.ai/the-batch/falling-llm-token-prices-and-what-they-mean-for-ai-companies/
