The AI market looks paradoxical. Free products become more capable every year, while the price of maximum access rises from $20 to $100, $200 and now $500 a month. Both trends come from the same technological progress. Yesterday’s frontier becomes cheaper to reproduce; today’s frontier spends more computation on work that was previously impossible.
Stanford’s 2025 AI Index gives the diffusion curve concrete numbers. The cost of querying a model at roughly GPT-3.5 benchmark quality fell from about $20 to $0.07 per million tokens between November 2022 and October 2024—more than a 280-fold decline. The smallest model crossing a defined MMLU threshold shrank from 540 billion parameters in 2022 to 3.8 billion in 2024, a 142-fold reduction.[1]
The two curves already visible in the market
| Measure | Value | Context |
|---|---|---|
| Fixed-capability inference cost | $20 to $0.07 per million tokens | Stanford’s comparison at approximately GPT-3.5 benchmark quality.Source 1 |
| Model size above a fixed MMLU threshold | 540B to 3.8B parameters | A 142-fold reduction in the smallest model crossing the reported threshold.Source 1 |
| Hardware performance trend | About +43% a year | Stanford’s reported annual improvement.Source 1 |
| Hardware price trend | About −30% a year | Reported annual cost decline for AI hardware.Source 1 |
| Hardware energy efficiency | About +40% a year | Reported annual improvement.Source 1 |
| Low-cost consumer tier | ChatGPT Go at $8 in the U.S. launch | Designed for broader access to everyday capabilities.Source 2 |
| Highest personal capacity tier | ChatGPT Pro 500 at $500 | Highest included usage and Astra Ultrafast among the current Pro tiers.Source 3Source 4 |
The floor rises while the ceiling moves away
A cheap model can inherit years of research through distillation, better data, improved training and efficient architecture. It does not need to recreate the original system’s size to perform everyday writing, extraction, tutoring or coding assistance. This raises the minimum capability available to free and low-cost users.
The frontier simultaneously moves into harder territory: long agent runs, computer use, high-stakes research, large codebases, difficult mathematics and low-latency professional service. These workloads can consume far more output tokens, tools and reserved capacity. OpenAI’s current enterprise rate card spans GPT-6 Luna at $0.10 input and $0.50 output per million tokens to GPT-6 Astra at $10 and $50 respectively—a 100-fold rate difference before tool and service multipliers.[5]
How a frontier capability moves toward the mass market
| Stage | What happens | What remains scarce |
|---|---|---|
| Research frontier | A costly system demonstrates a new capability | Reliable access, speed and reproducibility |
| Product frontier | The capability is wrapped in tools, safety and a usable interface | Capacity, support and stable operation |
| Routing | Only difficult requests reach the most costly system | Correct task classification and fallback |
| Distillation and specialisation | Smaller systems learn a narrower or cheaper version | The full breadth of the original frontier |
| Hardware and serving optimisation | The same work uses fewer chips, watts or seconds | Capital, power and peak demand |
| Broad paid tier | Capability appears with moderate allowances | Maximum effort, long runs and priority latency |
| Free or bundled tier | Common tasks become widely accessible | Heavy use, newest features and exceptional workloads |
A $500 plan is selling scarcity as much as intelligence
The highest tier does not merely contain better words. It sells more included usage, priority and Astra Ultrafast. The user pays to run an expensive mode more often and with lower latency. A light consumer sees little value because ordinary tasks already work on cheaper systems. A professional with repeated high-value tasks may be buying capacity, continuity and time rather than a simple model upgrade.[3][4]
This is why comparing only model names can mislead. A future $20 product may outperform today’s $500 product on many tasks because the technology diffuses. The future premium tier will then be attached to new, harder work. Price convergence happens at a fixed capability level, not necessarily at the moving frontier.
The access question is which curve matters for opportunity
Mass welfare improves when education, research, file analysis and everyday problem-solving move down the cost curve quickly. Inequality can still widen if the frontier tier supplies commercially decisive capabilities—faster agents, more reliable research or greater volume—that only firms and wealthy users can afford. The answer is not to freeze the frontier. It is to evaluate whether the lower curve remains capable enough for meaningful participation.
This dual curve is likely to define consumer AI for years. Free systems will feel astonishing compared with what existed a short time earlier. Professionals will continue paying high prices for the next increment of capability, allowance and speed. The mass-market achievement is not eliminating premium AI. It is ensuring that the cheaper curve rises fast enough to make powerful everyday intelligence broadly useful.
Sources & further reading
- Research and Development: 2025 AI Index Report Stanford HAI · Accessed 10 Oct 2026
- Introducing ChatGPT Go, now available worldwide OpenAI · 16 Jan 2026 · Accessed 10 Oct 2026
- About ChatGPT Pro tiers OpenAI Help Center · Accessed 10 Oct 2026
- ChatGPT release notes OpenAI Help Center · Accessed 10 Oct 2026
- ChatGPT Rate Card: enterprise token-based pricing OpenAI Help Center · Accessed 10 Oct 2026