The cost of AI intelligence has fallen through the floor since GPT-4’s release – with massive implications for how businesses can leverage these technologies.
I’ve been tracking price and performance metrics across language models, image generation, and video creation. The numbers are striking – we’re seeing exponential improvements in cost efficiency while quality keeps climbing.
## Language Models: 60× Cheaper in Just a Year
Let’s look at what’s happened with large language models since 2023:
– **GPT-3.5 Turbo (2023)**: ~70% MMLU score at $2/million tokens. That was already a 10× drop from GPT-3’s ~$20/million price tag.
– **GPT-4 (2023)**: 86% MMLU but priced at $30/million input tokens and $60/million for output (roughly $45/million for typical usage). GPT-4 Turbo later cut this by 2-3×.
– **GPT-4o Mini (2024)**: ~82% MMLU performance but costs only $0.15/million for input and $0.60/million for output. That’s approximately 60× cheaper than the original GPT-4 while delivering comparable intelligence.
– **DeepSeek R1 (2024)**: Matches GPT-4-level reasoning at just $2.19/million tokens – about 25× cheaper than comparable OpenAI models.
The trend is clear: LLM costs dropped more than 85% in 2023 alone, and the pace has only accelerated in 2024.
## Image Generation: From Cents to Fractions of a Penny
Image AI has seen even more dramatic price compression:
– **MidJourney (v5, 2022-2023)**: ~$0.03 per image through subscription pricing.
– **Stable Diffusion services (2024)**: Now offering 512×512 images for ~$0.0006 each (about $0.0023 per megapixel).
– **FLUX 1.0 “Schnell” (2024)**: $0.0014 per image – 20× cheaper than MidJourney was just months ago, while requiring only 4 diffusion steps.
Image generation costs have plummeted 10-100× in two years while quality has improved from “obviously AI-generated” to near-photorealism in many cases.
## Video Generation: From Impossible to Affordable
AI video generation has undergone the most dramatic transformation:
– **Google Veo 2 (2025)**: 4K resolution at $0.50/second ($30/minute).
– **Alibaba Wan 2.1 (2024)**: Open-source model that costs approximately $0.04/second on cloud GPUs or nearly zero on local hardware.
– **LTX-V (2024)**: Free, open-source real-time generator where the only cost is your local GPU runtime.
Video generation has gone from prohibitively expensive to pennies per second, with quality jumping from primitive 256×256 animations to usable 480p/720p with reasonable coherence.
## The Efficiency Explosion
Across all modalities, we’re seeing the same pattern:
1. **Costs drop dramatically**: 80-90% year-over-year decreases
2. **Quality improves simultaneously**: Higher resolutions, better coherence, more intelligence
3. **Speed increases**: Faster inference times enable new use cases
For example, GPT-4o mini offers ~82% MMLU performance for less than $1/million tokens compared to GPT-3.5’s ~70% for $2+/million tokens just a year prior. That’s a 3× improvement in quality-per-dollar in a single year.
## Jevons’ Paradox: Why Usage Will Explode
Classic economics predicts this situation perfectly. [Jevons’ Paradox](https://en.wikipedia.org/wiki/Jevons_paradox) states that when a resource becomes more efficient to use, people don’t use less of it – they use more.
As AI gets cheaper:
1. **Adoption accelerates**: Companies integrate AI into daily operations because the ROI is undeniable.
2. **Consumption patterns change**: Users run dozens of cheaper model calls instead of carefully rationing expensive ones. OpenAI has noted that cheaper models enable chaining multiple specialized calls together.
3. **New applications emerge**: Real-time, always-on AI becomes feasible in software, IoT devices, games, and media creation.
I’m already seeing this in my consulting work. Clients who balked at spending $100/month on AI tools are now running thousands of queries daily because the incremental cost has become negligible.
## DeepSeek: The Price War Catalyst
One company deserves special attention for accelerating this trend: DeepSeek.
Their approach has been revolutionary in several ways:
– **Ultra-low pricing**: DeepSeek R1 at $2.19/million tokens (25× cheaper than OpenAI’s comparable offerings)
– **High margins despite low prices**: They’ve claimed a 545% profit margin ($87K daily cost vs $562K potential revenue) through clever optimizations of their [inference stack](https://adam.holter.com/deepseeks-secret-weapon-how-its-inference-stack-crushes-openai-at-545-profit-margins/)
– **Open approach**: Released model weights and detailed technical reports publicly
– **Market impact**: Triggered an industry-wide price war, pushing OpenAI to release GPT-4o mini as a competitive response
DeepSeek proved what many suspected: frontier AI can be both state-of-the-art and affordable. They trained a 670B parameter MoE model on a modest ~$5-6M budget and demonstrated that serving it could be highly profitable even at radically lower prices.
## What This Means For You
The plummeting cost of AI intelligence means several things:
1. **Experimentation is cheap**: Try more use cases because the failure cost is minimal
2. **Scale matters less**: Small businesses can access the same AI capabilities as enterprises
3. **Integration beats optimization**: Focus on integrating AI broadly rather than squeezing every token
4. **Differentiation shifts**: As raw AI becomes commoditized, the value moves to customization and workflow integration
We’re rapidly approaching a world where AI becomes baseline infrastructure – as ubiquitous and unremarkable as electricity or internet connectivity. The winners won’t be those who have access to AI (everyone will), but those who integrate it most effectively into their workflows.
The exponential improvement curve isn’t slowing down. If you’re planning AI strategy based on today’s costs, you’re already planning for an outdated world. What seems expensive today will likely be 80-90% cheaper within 12-18 months.
My advice? Start experimenting broadly now. The cost of waiting far exceeds the cost of the AI itself.
