OpenAI Cuts GPT-3.5 Prices 50%, Launches Embedding Models That Trade Accuracy for Cost

OpenAI just reshuffled its model lineup in a way that matters more for developers’ AWS bills than it does for splashy headlines. The company is rolling out two new embedding models, a refreshed GPT-4 Turbo preview, and a cheaper GPT-3.5 Turbo — while quietly making a policy change that data sent to its API won’t be used for training by default. That last bit is the kind of bureaucratic-sounding shift that actually moves the needle on enterprise adoption, since it directly addresses the privacy concerns that have kept some companies on the sidelines.

On the embedding side, the headline numbers are stark. The new text-embedding-3-small model hits a 44.0% average on the MIRACL multilingual benchmark, a serious jump from the 31.4% that the outgoing text-embedding-ada-002 managed. Even better for anyone watching their cloud spend: pricing dropped 5x to $0.00002 per 1,000 tokens. The larger text-embedding-3-large pushes performance further — 54.9% on MIRACL and 64.6% on MTEB — at $0.00013 per 1,000 tokens. But the real trick here isn’t just better benchmarks. OpenAI trained these models so developers can deliberately shorten the embedding vectors by passing a dimensions parameter, trading off some accuracy for dramatically smaller storage footprints. A 256-dimension slice of the large model still outperforms the full 1536-dimension output of its predecessor. For anyone running vector databases at scale, that flexibility is the kind of thing that changes architecture decisions.

GPT-3.5 Turbo gets its third price cut in a year, with input costs halved to $0.0005 per 1,000 tokens and output down 25% to $0.0015. The new gpt-3.5-turbo-0125 model also fixes a text encoding bug that borked non-English function calls and promises better accuracy when responding in requested formats — a practical improvement for developers building structured data pipelines. Meanwhile, the GPT-4 Turbo preview update, gpt-4-0125-preview, takes aim at the “laziness” problem where the model would sometimes just… stop completing tasks. Over 70% of GPT-4 API requests have already moved to Turbo, drawn by the 128k context window and lower pricing.

The subtext across all these updates is that OpenAI is fighting a multi-front war. On one side, open-source models keep getting better and cheaper. On another, competitors like Anthropic and Google are pushing hard on enterprise features. Cutting prices while giving developers knobs to dial between performance and cost isn’t a revolutionary move — it’s table stakes now. But adding native embedding shortening and a no-training-by-default data policy suggests OpenAI understands that the next phase of API adoption hinges on cost control and compliance, not just benchmark scores.