Why AI Costs Keep Rising (Even As Prices Fall)

Описание: AI got 100x cheaper. So why is your bill higher? The price of one token fell about 300x since GPT-4, but a task now burns far more tokens (hidden thinking, agent loops) and the best models still cost near GPT-4 prices. Here is why, and what to do about it.
HYPE reads the headline; AUDIT opens the price pages. We start the Cost of Intelligence Index: the price of ONE solved task at a fixed accuracy. Everything is as of Oct 13, 2026.
Subscribe (a new drawing every few days):    / @costofintelligence  

0:00 AI got 100x cheaper. So why is the bill higher?
0:26 The collapse: price per token since GPT-4
1:42 Reason 1: hidden thinking tokens
2:42 Reason 2: agents resend the whole context
3:32 Reason 3: the top tier never got cheaper
4:16 Reason 4: plans with limits
4:51 Reason 5: Jevons paradox, cheaper means more use
5:47 The Cost of Intelligence Index: three tasks
8:16 Your 3 moves: consumer, builder, business
9:25 The paradox in one line

OUR MATH
Blended price = (3 x input + 1 x output) / 4, as Epoch AI does: GPT-4 ($30 in, $60 out) = $37.50 per 1M tokens; GPT-6 Luna ($0.10, $0.50) = $0.20; 37.50 / 0.20 = 188x cheaper (input alone 300x). The top tier, GPT-6 Astra ($10, $50) = $20.00, so only 1.9x cheaper than GPT-4. Epoch: $37.50 to $0.13 for GPT-4-level science questions = 288x.
Agent loop (our assumptions: 5,000 starting tokens, the model writes 500 and a tool adds 1,000 per turn, 20 turns, GPT-6.1 Sol prices): the model reads 5,000 + 1,500 x (turn - 1) tokens each turn = 385,000 tokens billed, while the last context is only 35,000 (11x). Cost: 385,000 x $2 / 1M + 10,000 x $10 / 1M = $0.87; with cached input at $0.10 and $2.50 cache write = $0.22.
Same price tag, different bill: Claude Sonnet 5.5 and GPT-6.1 Sol both list at $2 / $10, but one task costs $5.46 vs $0.72 (Artificial Analysis), 5.46 / 0.72 = 7.5x. Hidden thinking: 146,633 / 197,430 = 74% of Sonnet's output tokens.
Index tasks (price per solved task, ARC Prize at list prices): ARC-AGI-1 at 75% or more: $26 to $0.012 = about 2,000x. ARC-AGI-2 at 50% or more: $1.90 to $0.042 = 45x. ARC-AGI-2 at 80% or more: $13.62 to $0.10 = 136x. Consumer: Opus 5.5 $5.98 vs Luna $0.068 per task = 88x. Builder: GPT-6.1 Sol Max costs 2.5x Medium for 7.5 more points. Caching: $0.10 vs $2.00 = 95% off.
Token counts per run are not published for 2025-26 ARC entries; the Index is our own simple stand-in, not an official benchmark.

More drawings:
• Is AI a Bubble? Follow the Money:    • Is AI a Bubble? Follow the Money (Explaine...  
• GPU vs TPU: Can Anyone Break Nvidia's Grip?:    • GPU vs TPU: Can Anyone Break Nvidia's Grip?  

SOURCES (all as of 13 Oct 2026)
• OpenAI API pricing: https://developers.openai.com/api/doc...
• OpenAI, reasoning models guide (reasoning tokens are billed as output): https://developers.openai.com/api/doc...
• OpenAI Help Center, About ChatGPT Pro tiers: https://help.openai.com/en/articles/9...
• Anthropic, Claude API pricing: https://platform.claude.com/docs/en/a...
• Anthropic, Claude plans: https://claude.com/pricing
• Google, Gemini subscriptions: https://gemini.google/us/subscriptions/
• Artificial Analysis, models (cost per task, tokens per task): https://artificialanalysis.ai/models
• ARC Prize leaderboard data, ARC-AGI-1 (generated 5 Oct 2026): https://arcprize.org/media/data/leade...
• ARC Prize leaderboard data, ARC-AGI-2 (generated 5 Oct 2026): https://arcprize.org/media/data/leade...
• ARC Prize, o3 preview results (20 Dec 2024): https://arcprize.org/blog/oai-o3-pub-...
• Epoch AI, LLM inference price trends: https://epoch.ai/data-insights/llm-in...
• Stanford HAI, AI Index Report 2025 (280-fold drop, Nov 2022 to Oct 2024): https://hai.stanford.edu/ai-index/202...
• Alphabet Q2 2026 earnings remarks (22 billion tokens per minute; a company claim): https://blog.google/company-news/insi...
• Menlo Ventures, State of Generative AI in the Enterprise (Dec 2025; an investor's estimate): https://menlovc.com/perspective/2025-...
• W. S. Jevons, The Coal Question (1865), as cited: https://en.wikipedia.org/wiki/Jevons_...

Simplified for clarity. Prices change often; these are list prices on 13 Oct 2026. Company claims are labelled as such. Not financial or investment advice.
Music: "Investigations" by Kevin MacLeod (incompetech.com), licensed under Creative Commons: By Attribution 4.0 License, http://creativecommons.org/licenses/b...
Voices are AI-generated. HYPE and AUDIT are fictional characters.