OpenAI cuts GPT-5.6 Sol API pricing through at least November 21, 2026

Subscribers do not get this discount. They already receive subsidized compute relative to the $20 to $200 subscription fee, and heavy individual use would likely become consumption-priced otherwise.

I love the price war, and I hope open-source models keep pushing it. But open weights are not open source: current models are mostly opaque binaries that can run locally.

Anthropic's answer will probably be to slow its models down further. I do not think it is keeping pace in this race.

I read this as a 20% input and 33% output discount for Sol through at least November 21, 2026. Sol still costs 20 times Luna, but it is much more competitive with Anthropic and other providers.

These expiring-token promotions feel like a softer version of Anthropic's "use your free tokens before they expire" campaigns. It feels ominous, as if providers have reached the top of the demand curve and need lower prices for further growth.

I use Codex every day, but I want providers to call models small, medium, and large. Remembering what Sol means and why it beats another model is already too much cognitive overhead, which feels like commoditization.

The Sol, Terra, Luna names are basically large, medium, small with less boring labels. I do not think any naming scheme will satisfy everyone, because "large" does not tell the average ChatGPT user why it is better.

Model names have been bad because newer lite models can beat older pro models. GPT-5.6's Luna and Sol often sit on the price-performance frontier, while Terra can disappear, though a larger model may still have world knowledge Luna lacks.

I would not call frontier labs greedy when they are losing enormous amounts of money. The investors and data-center backers expect to capture the remaining middle-class livelihoods, and something in that model has to give.

I am not sure Chinese models and inference hardware can be trusted for corporate, government, or military use. A recent HN post showed a model can be trained to change behavior on a specific day, and I worry about data calling home.

I expected a private company that invented AI to have a huge, durable moat. Instead, distillation, enough text input and output, and a similar transformer architecture make useful models reproducible, so intelligence may become a race to the bottom.

I think hardware deployment scale will be the future moat. A company with an order of magnitude more silicon could retain a grip on frontier models and inference demand; China or SpaceX look like possible candidates in five years.

I keep seeing claims that models are easily distilled and replicated, but where is the actual evidence? I have not seen it.

Model-to-model teaching was known in ML before LLMs. It may not appear in investor materials, but it has played out; US-provider requirements and corporate onboarding friction still create some moat.

I expect two tiers: expensive "great AI" and cheap, good-enough AI. Chinese companies can make AI too, which should push costs down, while specialized models do not need frontier-level expense.