AI has gotten cheaper more quickly than any other transformative technology in history. The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year. That price drop is four times faster than DNA sequencing, six times faster than compute, 18 times faster than lithium batteries, and (in the century up to 1973) 54 times faster than electricity.
[...]
The next chart shows some examples. On January 31, 2025, OpenAI released a new iteration in its series of “reasoning” models, called o3. We estimate that for an average cost of 30 cents per question, it could achieve a 75% score on GPQA Diamond, a multiple-choice exam covering PhD-level physics, chemistry, and biology.1 Just under 18 months later, OpenAI released GPT-5.6 Luna. It scored just as well — for four hundredths of a penny per question ($0.0004). That is a 725-fold drop in the price of thought in under 18 months. It is like the sticker price on a new car falling from $50,000 to $69. No other general-purpose technology in history appears to have gotten so cheap so fast.
[...]
Our analysis comes with major caveats. AI companies may be expressly training their models for some benchmarks (“benchmaxxing”), so that improvement on the benchmarks outstrips improvement for real-world tasks. Even if they are not, doing well on a benchmark is not synonymous with useful work. Because we focus on the frontier — the absolute cheapest model capable of any given level of performance — we implicitly posit an AI user who relentlessly searches for the most cost-effective model for each task, when real users do not switch models so often, and therefore do not reap quite the same savings. Our data are incomplete and noisy: the timeframe is barely three years, and we do not include all combinations of AI model and benchmark. Prices drop differently for different models, benchmarks, time periods, and performance ranges, and there are many reasonable ways to average over this variegated experience. Overall, while we believe that our bottom-line numbers are reasonably representative of reality, they should not be read as exact.
I love how good local models are and how optimized the software is. I was gifted a couple of old workstation cards (Turing generation RTX 6000s) and put them to use hosting Qwen 3.8 27B. With...
I love how good local models are and how optimized the software is. I was gifted a couple of old workstation cards (Turing generation RTX 6000s) and put them to use hosting Qwen 3.8 27B. With llama.cpp I can split the weights across the cards so they run in parallel. Multiple requests can batch together. I get 60 tok/s for one request or 40 tok/s each for two concurrent requests. And the model’s way smarter than it should be given the size. In a few years new mid-range consumer hardware will actually be hosting models that are worth running.
For coding, currently I talk to Astra to make a plan and tell it to use a Luna subagent to implement each commit. Luna is extremely cheap and I don't worry that using Luna alone will ever exhaust...
For coding, currently I talk to Astra to make a plan and tell it to use a Luna subagent to implement each commit. Luna is extremely cheap and I don't worry that using Luna alone will ever exhaust my $20/month plan. Astra is too expensive for coding use unless you're using it professionally, but I think the planning goes better.
And I haven't even switched to the latest version of Luna yet, which will be half the price.1 It doesn't seem to be available via a ChatGPT subscription yet.
1 Yeah, it's an "introductory" price but I think that's OpenAI hedging their bets in case there's a capacity crunch. These price drops have a tendency to become permanent due to competition.
I've been doing a lot of work where I have a GLM 5.3 sub-agent launched for every file in a repo (for small projects) and then hand over findings to Fable or Opus to prune.
I've been doing a lot of work where I have a GLM 5.3 sub-agent launched for every file in a repo (for small projects) and then hand over findings to Fable or Opus to prune.
I've been using Deepseek Flash 4.1 served through fireworks.ai for report development in OpenChamber and its remarkable. Cost-wise, assembling a 45-page report has set me back $2.19. Granted, I'm...
I've been using Deepseek Flash 4.1 served through fireworks.ai for report development in OpenChamber and its remarkable. Cost-wise, assembling a 45-page report has set me back $2.19. Granted, I'm making heavy use of skills, tool calls, and retrieval augmented generation, but DS4.1 is really good, with Fireworks raising prices on October 1st to deal with "unprecedented demand". After I'm done, I might run it through Opus 5.5, as the word is that it writes well, but the total cost is will still be very low. I'm excited to see what lower cost open weights models are going to be capable of in the coming months.
From the article:
[...]
[...]
I love how good local models are and how optimized the software is. I was gifted a couple of old workstation cards (Turing generation RTX 6000s) and put them to use hosting Qwen 3.8 27B. With llama.cpp I can split the weights across the cards so they run in parallel. Multiple requests can batch together. I get 60 tok/s for one request or 40 tok/s each for two concurrent requests. And the model’s way smarter than it should be given the size. In a few years new mid-range consumer hardware will actually be hosting models that are worth running.
For coding, currently I talk to Astra to make a plan and tell it to use a Luna subagent to implement each commit. Luna is extremely cheap and I don't worry that using Luna alone will ever exhaust my $20/month plan. Astra is too expensive for coding use unless you're using it professionally, but I think the planning goes better.
And I haven't even switched to the latest version of Luna yet, which will be half the price.1 It doesn't seem to be available via a ChatGPT subscription yet.
1 Yeah, it's an "introductory" price but I think that's OpenAI hedging their bets in case there's a capacity crunch. These price drops have a tendency to become permanent due to competition.
I've been doing a lot of work where I have a GLM 5.3 sub-agent launched for every file in a repo (for small projects) and then hand over findings to Fable or Opus to prune.
I've been using Deepseek Flash 4.1 served through fireworks.ai for report development in OpenChamber and its remarkable. Cost-wise, assembling a 45-page report has set me back $2.19. Granted, I'm making heavy use of skills, tool calls, and retrieval augmented generation, but DS4.1 is really good, with Fireworks raising prices on October 1st to deal with "unprecedented demand". After I'm done, I might run it through Opus 5.5, as the word is that it writes well, but the total cost is will still be very low. I'm excited to see what lower cost open weights models are going to be capable of in the coming months.