In Today’s Issue:

📄 What the primary documents actually say, including three things almost every secondary account got wrong

📜 What Strauss really promised in 1954, and why the correct reading cuts against both sides of this argument

📉 Whether the cost of a fixed capability is genuinely collapsing, and how thin the evidence under the famous numbers turns out to be

💻 Why a large part of the celebrated hardware curve is a number format rather than a chip

⚡ Where the meter ended up: enterprise consumption billing, memory contracts, and an electricity market that has hit its own price ceiling three times running

Dear Readers,

On the thirtieth of July 2026, OpenAI published a price list that moved in three directions at once. GPT-5.6 Luna, the small model at the bottom of the ladder, fell 80 percent, from one dollar to twenty cents per million input tokens. Terra, the middle tier, fell 20 percent. And Sol, the model at the top, did not move at all: "Sol pricing remains unchanged" (OpenAI, 07/30/2026). The same post kept a premium attached to speed, replacing Priority Processing with a Fast mode that serves Sol at "up to 2.5x faster speeds than Standard processing at twice the price, with no change in intelligence".

Eight days later the picture had filled in. OpenAI dropped the message quota for its free users. DeepSeek, one of the cheapest large API providers on the market, had already been charging double during peak hours since the end of June, and on 6 August it told customers that a significant general increase was on the way (South China Morning Post, 06/30/2026; DeepSeek pricing page, 08/06/2026). Anthropic, meanwhile, has a mid-tier model scheduled to get 50 percent more expensive on 1 September (Anthropic pricing documentation, retrieved 08/08/2026). In the week everyone read as the arrival of nearly free intelligence, the floor of the market was busy warning people it was going up.

The phrase that attached itself to those eight days is seventy years old. In 1954 the chairman of the US Atomic Energy Commission, Lewis Strauss, told a room of science writers that their children would enjoy electrical energy too cheap to meter. It became the most famous broken promise in the history of technology forecasting, and it is now being attached to language models by people who mean it as praise. Sam Altman has written that intelligence too cheap to meter is well within grasp ("The Gentle Singularity", 2025), and a customer testimonial hosted inside OpenAI's own price-cut post calls Luna the closest we have come to exactly that (OpenAI, 07/30/2026).

Almost nobody goes back to check what Strauss actually meant. His claim was about billing rather than about price, and he said so himself a few days later. Once you take his definition seriously, the interesting question stops being whether intelligence is getting cheap. Prices are plainly falling, and that version of the question was settled long ago. The question becomes which tier of this industry gets which billing architecture, who assigns it, and what the assignment tells you about where the scarcity actually sits. So: is intelligence really becoming too cheap to meter, or is the meter simply being moved to places where nobody is reading it?

All the best,

Kim Isenberg

Start with what is on the record, because almost every secondary account of this week gets at least one element wrong.

On 9 July 2026 OpenAI published list prices for the GPT-5.6 family: Sol at 5 dollars input and 30 dollars output per million tokens, Terra at 2.50 and 15, Luna at 1 and 6 (OpenAI, 07/09/2026). Three weeks later came the cut. Terra to 2 and 12, Luna to 0.20 and 1.20, Sol unchanged (OpenAI, 07/30/2026). Eighty percent at the bottom, twenty in the middle, zero at the top.

One detail here was worth checking against OpenAI's own documentation, and it checks out. GPT-5.5, the previous generation's frontier model, is listed at 5 dollars input and 30 dollars output per million tokens (OpenAI API documentation, retrieved 08/08/2026). The price of frontier intelligence at OpenAI has now held steady across two model generations while the floor beneath it fell by four fifths.

A correction has to be made immediately, because the tempting version of this story is wrong. Frontier prices do fall, just not on this list on this day. Anthropic's own price list puts Claude Opus 5 at 5 dollars input and 25 dollars output per million tokens against 10 and 50 for Fable 5, and runs its own research-preview Fast mode at exactly double the standard rate (Anthropic pricing documentation, retrieved 08/08/2026). A frontier-class model at half the price of its own flagship, landing at Sol's input price and below Sol's output price, and a second vendor pricing speed at 2x. The claim here is therefore narrow: inside one company's list, on one day, the cut ran 80 percent at the bottom and zero at the top.

The same list carries the sharper finding. Claude Sonnet 5 sits on introductory pricing of 2 and 10 dollars through 31 August 2026, after which it rises to 3 and 15. A scheduled 50 percent increase on a mid-tier frontier model, published by the vendor itself, sits in the middle of the cheapest week anyone can remember.

The day before its cut, OpenAI published its own explanation: production kernels rewritten autonomously by GPT-5.6 Sol lowered end-to-end serving costs by 20 percent, and a better speculative decoding draft model raised token generation efficiency by more than 15 percent (OpenAI, 07/29/2026). Those are vendor claims against an internal baseline, with no external verification available. The same post describes OpenAI's operating environment as "a compute-constrained world where model demand is growing faster than capacity" (OpenAI, 07/29/2026). A company that calls itself capacity-constrained is not a company that thinks its input has become abundant.

On 6 August the consumer half landed. Luna became the default model for Free and Go users, and those users would get unlimited text chats and a Think button for harder questions, subject to abuse guardrails, with limits still applying to file uploads, images and other tools (OpenAI, 08/06/2026). ChatGPT serves a billion people weekly, according to the same post, though the unlimited tier reaches only the Free and Go users inside that number. It covers text only, and as of publication it had been announced rather than rolled out.

Now the part most coverage inverts. DeepSeek did not cut prices during this window, and its move toward finer metering began well before it. Peak-hour pricing at double the standard rate, applying from 09:00 to 12:00 and 14:00 to 18:00 Beijing time, went out to customers by email on 30 June 2026, a full month before OpenAI's cut (South China Morning Post, 06/30/2026). The base prices stayed put throughout: archived snapshots show V4-Flash at 0.14 dollars cache-miss input and 0.28 output on 1 May, 2 July and 28 July 2026 (DeepSeek pricing page, archived snapshots, 05/01/2026 to 07/28/2026). What happened inside the window came on 6 August, when a notice appeared saying the company plans to raise overall pricing in the near future, with a significant increase expected (DeepSeek pricing page, 08/06/2026). No new number has been published, so this is a warning rather than an increase. The changelog for the 31 July release adds that the new V4-Flash-0731 keeps the same architecture and size as its predecessor and "was only re-post-trained" (DeepSeek changelog, 07/31/2026).

That sequence is the sharpest data point available for a piece about intelligence becoming too cheap to meter, and it is sharper than the reaction story it replaced. The cheapest end of the market has been metering more finely since June, and it spent the first week of August telling customers to expect worse. The reported motive, as far as journalism has established it, is a data centre in Inner Mongolia and a plan to secure a gigawatt of computing power, for which the company needs to build up funding (Bloomberg, 08/06/2026). A company raising prices to finance expansion is telling you demand is strong, not that the technology is uneconomic.

One housekeeping note: the 20,000-GPU figure attached to DeepSeek's increase comes from a Bloomberg report about Moonshot, a different company (wccftech, 08/06/2026).