The price of intelligence: tracking AI API costs since 2020

I've been tracking AI API prices since GPT-3's beta in 2020. Today I'm publishing the full timeline.

47 data points. 12 providers. 5 years. The pattern is unlike anything I've seen in technology.

The complete timeline (output token pricing, best available model)

| Date | Model | Output/M tokens | Quality (MMLU equiv.) | |------|-------|-----------------|----------------------| | Jun 2020 | GPT-3 Davinci | $60.00 | ~43% | | Mar 2021 | GPT-3 Davinci (price cut) | $40.00 | ~43% | | Jan 2022 | GPT-3 Davinci (instruct) | $20.00 | ~50% | | Dec 2022 | ChatGPT (gpt-3.5-turbo) | $2.00 | ~70% | | Mar 2023 | GPT-4 | $60.00 | ~86% | | Jun 2023 | GPT-3.5-turbo (updated) | $1.50 | ~70% | | Nov 2023 | GPT-4 Turbo | $30.00 | ~86% | | Mar 2024 | Claude 3 Haiku | $1.25 | ~75% | | May 2024 | GPT-4o | $15.00 | ~89% | | Jun 2024 | Claude 3.5 Sonnet | $15.00 | ~89% | | Jul 2024 | GPT-4o mini | $0.60 | ~82% | | Dec 2024 | Gemini 2.0 Flash | $0.40 | ~84% | | Jan 2025 | DeepSeek V3 | $1.10 | ~87% | | Apr 2025 | Gemini 2.5 Flash | $0.60 | ~86% | | Nov 2025 | Best "cheap" model | $0.30 | ~86% |

Sources: OpenAI, Anthropic, Google, Cohere, AI21 Labs, DeepSeek, Artificial Analysis, provider blogs and announcements.

The two curves

There are actually two stories in this data:

Story 1: Same-quality pricing fell 100x. GPT-3-level quality cost $60/M in 2020. Today, models at or above GPT-3 quality cost under $0.60/M.

Story 2: Best-available pricing fell, then rose, then fell again. When GPT-4 launched in March 2023, the price of "best available" jumped back to $60/M. Then it fell as GPT-4 got cheaper and competitors appeared. Today, the best available models range from $10-75/M output.

| Price metric | 2020 | 2023 (GPT-4 launch) | Today (Nov 2025) | |-------------|------|--------------------|--------------------| | Best model, output/M | $60.00 | $60.00 | $75.00 (Opus 4) | | GPT-3 equivalent quality, output/M | $60.00 | $1.50 | $0.30 | | Cheapest "usable" model, output/M | $60.00 | $1.50 | $0.30 |

The ceiling (best model) has stayed roughly flat at $60-75/M over five years. The floor (cheapest usable model) dropped 200x.

The deflation rate

| Period | Price drop (cheapest quality-equivalent) | Annualized rate | |--------|----------------------------------------|-----------------| | 2020-2021 | -33% | -33% | | 2021-2022 | -50% | -50% | | 2022-2023 | -93% (ChatGPT/GPT-3.5-turbo) | -93% | | 2023-2024 | -73% (GPT-4o mini) | -73% | | 2024-2025 | -50% (Gemini Flash) | -50% |

The fastest deflation was 2022-2023 when ChatGPT dropped the price of conversational AI from $20 to $1.50 per million tokens. A 93% price drop in one year.

The rate has slowed from 93% to 50% annually. Still massive deflation, but the curve is flattening as prices approach marginal compute costs.

Comparison to other technology cost curves

| Technology | Time to 100x cost reduction | Starting point | |-----------|---------------------------|---------------| | DNA sequencing | 13 years (2001-2014) | $100M per genome | | Solar panels | 20 years (2000-2020) | $4.50 per watt | | Hard drive storage | 15 years (1995-2010) | $0.10 per MB | | AI API pricing | 5 years (2020-2025) | $60 per M tokens |

Sources: IEA, National Human Genome Research Institute, industry data.

AI API pricing achieved 100x cost reduction in 5 years. DNA sequencing took 13 years. Solar panels took 20 years.

This is the fastest cost reduction curve I can find in any technology industry. The combination of hardware improvements (GPU generations), software improvements (MoE, quantization), and competitive pressure (DeepSeek, open source) created a perfect storm of deflation.

Where the floor might be

| Component | Estimated floor cost | Rationale | |-----------|---------------------|-----------| | Electricity per M tokens | $0.003-0.01 | Physics limit (Landauer + overhead) | | Hardware depreciation per M tokens | $0.02-0.05 | Assumes B200-class hardware | | Data center overhead | $0.01-0.03 | Cooling, networking, maintenance | | Margin | $0.01-0.05 | Provider profit | | Estimated floor | $0.05-0.15 | For commodity-tier models |

I think the floor for commodity-tier AI inference is around $0.05-0.15 per million output tokens. We're at $0.30 today. Another 2-6x reduction is plausible before we hit physical compute cost limits.

Premium models will always command a premium. But the "good enough for most tasks" tier is approaching a price floor where AI inference becomes essentially free for individual use.

Five years of tracking

This dataset started as a column in a spreadsheet I built to track my own API costs. Five years later, it's 47 rows with notes, dates, model names, and the occasional personal annotation ("Dec 2022: ChatGPT launched. My phone won't stop buzzing.").

The numbers tell the most important story in AI: intelligence is getting cheaper faster than any technology in history.

---

If you found this interesting, you might also like:

• The cost of running an AI startup in 2022: a data breakdown
• The real cost of training Llama 2: Meta's numbers vs my estimates
• Every LLM API price drop in the last 12 months, in one chart
• Every AI pricing change in January 2025, tracked
• The real cost of AI agents: I tracked token usage for 50 agentic tasks

-- dataku