The cost of intelligence is collapsing
GPT-4-class input pricing fell ~93% in two years: $30 → $2 per Mtok. As the reasoning core commoditises, value moves up the stack to the agent: orchestration, memory, tools and outcomes. The model is the cheapest part.
| subject | metric | value | as of | src |
|---|---|---|---|---|
| GitHub Copilot | pricing model | usage-based tiers (Pro $10 / Pro+ $39 / Max $100) with monthly AI credit allowances ($15/$70/$200); 1 AI credit = $0.01 | 2026-07-10 | ↗ T1 |
| CX agents (Intercom/HubSpot) | pricing model | per-resolution / outcome-based pricing | – | ↗ T1 |
| Salesforce Agentforce | pricing model | consumption-based Flex Credits: 20 credits = $0.10 / standard action, 30 credits = $0.15 / voice action; 100k credits pack = $500 | 2026-07-10 | ↗ T1 |
| Frontier model tokens | cost trend | input $/Mtok to reach a fixed capability falls ~40x/yr (GPT-4 quality); ~50x/yr median across benchmarks | – | ↗ T1↗Artific |
| Agent consumers | trend | shift to headless / API-first (no UI) consumption | – | ↗ T1 |
| OpenAI GPT-4 class (input) | price $/Mtok | $30.00 / Mtok (GPT-4 launch) | 2026-07-10 | ↗ T1 |
| OpenAI GPT-4 class (input) | price $/Mtok | $10.00 / Mtok (GPT-4 Turbo) | 2026-07-10 | ↗ T1 |
| OpenAI GPT-4 class (input) | price $/Mtok | $5.00 / Mtok (GPT-4o) | 2026-07-10 | ↗ T1 |
| OpenAI GPT-4 class (input) | price $/Mtok | $2.50 / Mtok (GPT-4o Aug) | 2026-07-10 | ↗ T1 |
| OpenAI GPT-4 class (input) | price $/Mtok | $2.00 / Mtok (GPT-4.1), −93% vs launch | 2026-07-10 | ↗ T1 |
| LLM inference price / fixed quality | annual decline (Epoch AI) | ~50x/yr median · ~200x/yr since Jan 2024 · ~40x/yr for GPT-4-class quality | – | ↗ T1 |
| Frontier flagship (GPT-5.5) | price change vs prior gen | input $2.50→$5.00 · output $15→$30 vs GPT-5: the frontier premium re-inflates even as commodity tokens keep falling | – | ↗ T1 |
GPT-4-class input pricing fell roughly 93% in two years, and value moved up the stack
Token prices have fallen consistently and by large multiples, faster than almost any comparable input in computing.
Show more
That decline changes what is worth automating. Work that was uneconomic at last year’s prices becomes viable at this year’s, which is why agent deployment tends to lag capability by roughly the time it takes prices to halve. The constraint moving to the front is human review, not inference.
How much has the price of AI models fallen?
GPT-4-class input pricing fell by roughly 93% in about two years, from near $30 per million tokens to around $2. The same capability keeps getting cheaper with each model generation, so today's mid-tier model matches last year's frontier one at a fraction of the price.
Why has AI got so much cheaper?
Four things at once: faster hardware, much more efficient serving software, smaller models that match older large ones, and hard price competition between providers. Each one compounds on the others.
How much does one million tokens cost?
It depends entirely on the model. Frontier models are the most expensive and small models cost a small fraction of that. Output tokens usually cost several times more than input tokens. Providers publish current rates on their own pricing pages (OpenAI) (Anthropic).
What is the difference between input and output token pricing?
Input tokens are what you send: the prompt, any documents, and the conversation so far. Output tokens are what the model writes back. Output costs more because it is generated one step at a time. For agents, long context often makes input the bigger line.
Will AI keep getting cheaper?
Per token, the trend has held through every model generation so far. Per task is a different question: newer models reason for longer and use far more tokens on the same job, so a cheaper token does not automatically mean a cheaper answer.