OpenAI reports applying GPT-5.6 Sol to make the software that serves its models cheaper to run. Its post credits Sol with improvements to production GPU kernels, the routines that perform model calculations on graphics processors, and says they produced 20% lower serving costs.
A separate result concerned speculative decodingA speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.Full definition →, where a smaller model proposes several output tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → for the main model to check together. OpenAI reports more than 15% better token-generation efficiency from the improved version, described in its engineering account. These are two separate reported results, so they should not be added into a single percentage saving. Related Intel coverage covers the same account.
Terms in this piece · Glossary
speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
