▶ YouTube Summaries

← All summaries
t3dotgg#▶ Watch

AI is getting cheaper per level of intelligence, but bills grow because usage grows

2026-10-12 Mon ⏱ 18 min t3dotgg

Theo answers Prime's tweet asking why AI got more expensive when Sam Altman promised 10x cheaper per year. His answer is that it did get cheaper, but token price is the wrong measure. Three things matter: price per token, tokens needed per task, and the cost of reaching a given level of intelligence. Top-tier models cost more per token, yet the cost of a fixed capability level has dropped far more than 10x per year (about 20x for GPT-5 level, by his estimate). Bills still rise because models do bigger, longer tasks end to end and people use them far more, a Jevons paradox. He admits his 2024 "AI won't keep improving" take was badly wrong.

Sponsor: Railway

  • Cloud for full stack apps: database, CDN, persistent servers, plus skills, CLI and MCP so agents can deploy and manage apps.
  • Theo's Lakebed cloud runs on it, with production, staging and per-PR environments provisioned by agents. Estimated cost for his whole internal video pipeline and apps is under $50 per month.

The puzzle

  • Altman (July 2025) said intelligence too cheap to measure is the next frontier. Theo's own usage went from about $800-1,000 for all of 2025 to over $1,000 in tokens on a single day in 2026.
  • Newer models (Fable, Astra) cost more per token and also generate more tokens, so real-world AI bills keep climbing.
  • Prime's tweet pointed at the promised 10x yearly price drop versus Astra being far more expensive.

Dimension 1: price per token

  • Artificial Analysis: GPT-5 medium is $10 per million out and $1.25 per million in; GPT-6 Astra is $50 out and $10 in, so 5-10x up.
  • The cheap end also improved. GLM 5.3 Flash is $0.15 in and $0.50 out (often half off), about a 20-30th of GPT-5's price, and beats it in many agentic jobs.
  • Both ends moved: the top got pricier, the bottom got far cheaper.

Dimension 2: tokens per task

  • Gemini 3.5 Flash looks like less than half of GPT-5 and under a tenth of Astra and Fable per token. On cost per Intelligence Index task the gap shrinks from over 10x to under 3x, because it burns many tokens. Another Flash model reached over half the cost of GPT-5.6 Soul, a far pricier model.
  • Sonnet ended up more expensive than Astra on a per-task basis for the same reason.
  • Which cost matters depends on the job. For bulk classification of 800k input tokens, input price dominates. For agentic dev work, caching, reliability and token efficiency matter more.
  • On the cost versus intelligence chart (Pareto frontier), Astra holds many of the frontier spots and Fable is far to the right. Around a score of 45-46 a task costs $0.82; moving to the 50s costs $3.26, about 3x for a few points.

Dimension 3: cost at a fixed level of intelligence

  • Theo's preferred framing: fix a capability level and track its price. Matching benchmarks to GPT-5 was hard because old benchmarks saturated. DeepSeek V4 Flash is roughly equal to GPT-5 Codex high on the overlapping benches and wins some (Humanity's Last Exam, AI Omniscience).
  • V4 Flash is 21x cheaper on input and 55x cheaper on output. Even assuming it uses 3x more tokens, that is about 20x cheaper overall, beating the 10x goal.
  • Artificial Analysis tracking by level:
  • Under 10 points: $0.60 per task fell to $0.02 over about two years (Nov 2022 to Jul 2024).
  • 10-20 points: o1 Preview was about $15 per task in Sep 2024 with a score of 11. By Feb 2025 Grok 3 Mini scored four points higher for $0.16, about 99% cheaper in four months.
  • 20 points: o3 first broke it. GPT-5 Mini got the same score for $26 instead of $155, DeepSeek 3.2 for $10, Ling 3 Flash for $4.7.
  • 30s: started at $1.90 per task and soon fell much lower.
  • 40-50s: Opus 4.7 started at $3.85 per task (41 points). GPT-5.6 Terra then scored higher at almost half the price, GLM 5.3 higher again at under a third, and GLM 5.3 Flash nearly the same at a 100th.
  • Pattern: a new level launches at a modest premium, then prices fall fast, and drops are faster each time. The chart axes are logarithmic.

Why bills still grow

  • Usage changed: from autocompleting one function to taking a bug screenshot and landing the fix end to end. Models run longer, stay on task and need less intervention.
  • More tokens per prompt is success, not failure. This is Jevons paradox: cheaper, better steam power led to spending more on coal because it became usable everywhere.
  • Spend is tied to productivity. A $1,500 day would not have been possible with GPT-5, since it could not get that much work done.
  • Comparing GLM 5.3 Flash to GPT-5 on price alone undersells it. It is better on vision, tool calls and long-running agentic work, and with a search tool it would beat GPT-5 on knowledge tasks too. Only built-in knowledge is lower. Small models with good tools beat last year's frontier.
  • He expects Fable 5.1-level intelligence at GLM 5.3 Flash pricing to happen.

Takeaways

  • Theo owns up to his 2024 video "AI isn't going to keep improving", which he calls the worst take he has had. The benchmarks had stalled because they needed replacing.
  • "Intelligence too cheap to measure" has not arrived as Altman implied, but 10x cheaper per level was beaten by a wide margin.
  • Costs per level keep falling even if bills do not look like it.