GPT-3.5 cost $0.002/1K tokens in 2022. GPT-4o mini costs $0.00015 in 2024. The deflationary spiral is faster than Moore's Law and it's reshaping every AI business model built on markup pricing.
93%Price drop: GPT-class API tokens 2022–2024
$0.00015GPT-4o mini per 1K tokens (2024) vs $0.002 (2022)
Choose your depth. The data doesn't change — just the explanation.
When AI companies first started letting people use their AI through an app, it was expensive — like paying a lot per word to use a calculator. Now it's incredibly cheap. The price has dropped 93% in two years. This means every company that built a business by charging people more than it cost to use the AI is now in trouble — because the AI itself is almost free.
AI API prices have deflated faster than any computing commodity in history. The driving forces: GPU efficiency improvements (CUDA optimizations, FlashAttention), model architecture efficiency (smaller models achieving similar results), and intense competition among providers (OpenAI, Anthropic, Google, Groq, Together.ai). Any business built on AI markup pricing — charging users $X while paying OpenAI $Y — faces margin erosion as Y approaches zero. The implication: the money in AI must come from data, distribution, or proprietary models, not API arbitrage.
Price history: GPT-3 (Jan 2021): $0.0200/1K tokens. GPT-3.5-turbo (Mar 2023): $0.0020/1K. GPT-3.5-turbo-0125 (Jan 2024): $0.0005/1K. GPT-4o mini (Jul 2024): $0.00015/1K (input). This follows Wright's Law (cost decreasing at constant rate with cumulative production) at an unusually steep 60%/year decline. Competitive dynamics: Groq (LPU inference): $0.0001/1K. Fireworks.ai: $0.0002/1K. The floor is hardware cost + power: estimated at $0.00003-0.0001/1K at optimal utilization. Open-source self-hosted: ~$0.0004-0.0008/1K (including hardware amortization, TCO).
Price history source: artificialanalysis.ai (comprehensive API price tracking). OpenAI pricing changelog: platform.openai.com/docs/pricing. LLM price comparison: llmpricecheck.com, openrouter.ai/models. Compute cost floor: Epoch AI "Inference Cost Efficiency" analysis (2024). Wright's Law applied to AI inference: Hu et al. "Scaling Laws for Neural Language Models," NeurIPS 2020. Track: Together.ai, Groq, Fireworks.ai, Lepton.ai price history via Wayback Machine archive snapshots. Monitor: NVIDIA GPU spot pricing (lambda.ai/service/gpu-cloud), H100 hourly rates.
The Price Collapse
Every 6-12 months, the cost per AI API token has fallen dramatically. What cost $20 per million tokens in 2021 costs $0.15 in 2024.
OpenAI API Price per Million Tokens (USD) — Log Scale
OpenAI pricing history. Log scale because the linear drop is too steep to visualize otherwise.
Model
Launch Date
Price/1M Input Tokens
vs. GPT-3 Peak
GPT-3 (Davinci)
Jan 2021
$20.00
baseline
GPT-3.5-turbo
Mar 2023
$2.00
−90%
GPT-3.5-turbo-0125
Jan 2024
$0.50
−97.5%
GPT-4o mini
Jul 2024
$0.15
−99.25%
Gemini Flash 1.5
May 2024
$0.075
−99.6%
Llama 3 (self-hosted)
2024
~$0.40
−98%
Groq (Llama 3)
2024
$0.10
−99.5%
Price Compression Rate vs. Moore's Law
AI token prices falling faster than transistors doubled.
AI API Business Model Impact
Margin compression for AI middleware businesses.
The Middleware Massacre
Hundreds of startups built businesses on the premise "we call GPT-4, charge 3× markup, keep the difference." As API prices collapsed 93%, those markups vanished. The sustainable AI business model requires proprietary data, distribution moats, or vertical integration — not API arbitrage.
Sources
• OpenAI. Pricing history. platform.openai.com/docs/pricing.
• Artificial Analysis. API price tracking. artificialanalysis.ai.
• LLM Price Check. llmpricecheck.com.
• OpenRouter. Model pricing directory. openrouter.ai/models.