AI, Work and the Economy

Latest/Story

The falling price of AI

Four items, from 28 Sep 2026 to 18 Nov 2026, oldest first.

See also: Point 1 of our view: AI does more office work, cheaper

AI and technologyNew data

Epoch AI (28 Sep): the cost of reaching a fixed AI benchmark score falls 13x a year; 725x on GPQA in 18 months

Epoch AI finds the cheapest price of hitting a given score on five benchmarks has fallen 13x a year; o3's 75% on GPQA Diamond cost $0.30 a question in Jan 2025, matched 18 months later for $0.0004. That is 6x faster than compute and 18x faster than lithium batteries.

Why it matters. If the price of a given level of AI skill keeps falling this fast, the cost argument for keeping people on routine thinking tasks weakens quickly.

Next. Epoch's and METR's next measurements of frontier models.

Sources: epochai.substack.com

AI and technologyNew dataUpdate

OpenAI's GPT-6.1 Sol (29 Sep): near GPT-6 Astra on coding and computer use at $2/$10 per M tokens, 1/5 Astra's price

At DevDay on 29 Sep OpenAI released GPT-6.1 Sol, which it says nearly matches its flagship GPT-6 Astra ($10/$50 per million tokens) on agentic coding, computer use and professional work, at $2 input and $10 output. Factual errors fall from 11.4% to 7.7% against GPT-6 Sol, within 1.9 points of Astra.

Why it matters. Work that needed OpenAI's most expensive model three weeks ago can now be done, the company says, for a fifth of the price, which widens the range of jobs where an AI agent is cheaper than a person.

Next. Independent benchmark results over the coming week will show whether it really matches the flagship.

Sources: openai.com, techcrunch.com

AI and technologyNew dataUpdate

OpenAI's GPT-6.1 Terra (26 Oct): computer-use model at $1/$5 per million tokens, half GPT-6.1 Sol's price

OpenAI released GPT-6.1 Terra for ChatGPT Work, Codex and the API at $1 input and $5 output per million tokens, 50% below GPT-5.6 Terra and GPT-6.1 Sol. OpenAI reports it scores 15 points higher than GPT-5.6 Terra on Agents' Last Exam at about a third of the cost.

Why it matters. OpenAI's new model for operating computers on a worker's behalf costs half as much as its predecessor while scoring higher on the company's agent tests. If prices for such work keep halving every few weeks, the cost case for replacing routine office tasks may strengthen faster than firms and policy adjust.

Next. Independent benchmark results and rival pricing over the coming weeks.

Sources: openai.com, xda-developers.com · Point 1 of our view

AI and technologyNew dataUpdate

Microsoft (Ignite, 18 Nov): MAI Code 2 scores 96% of Claude Opus 5.5 on coding at about 5% of its price

Microsoft released seven in-house models; it says MAI Code 2 averages 98% of GPT-6.1 Sol and 96% of Claude Opus 5.5 across four coding benchmarks at about 40% and 5% of their prices, and that MAI Cyber 2 Flash matches GPT-6 Astra on a cyber benchmark at under 2% of its cost per task.

Why it matters. Microsoft says its new coding model comes close to the leading models at a small fraction of their price. If independent tests confirm it, the cost of automating software work is falling fast, while the labs that borrowed to build capacity may find prices falling faster than their revenue grows.

Next. Independent benchmark results on the new models.

Sources: microsoft.ai, news.ycombinator.com · Point 1 of our view