OpenAI drops new embeddings, fixes GPT-4 laziness

OpenAI just addressed two of developers' biggest daily frustrations: models cutting corners on tasks, and RAG pipelines getting expensive and heavy.

The headliner of the latest drop is a pair of next-generation embedding models: text-embedding-3-small and text-embedding-3-large. They replace the workhorse text-embedding-ada-002 model, bringing significant performance leaps alongside steep price cuts.

Smarter vectors you can actually shrink

The smaller text-embedding-3-small costs 5 times less than its predecessor, dropping to $0.00002 per 1k tokens. Despite being dirt cheap, its multi-language retrieval score on MIRACL jumped from 31.4% to 44.0%, and English MTEB ticked up to 62.3%.

The beefier text-embedding-3-large pushes dimensions up to 3072. It hits a 54.9% score on MIRACL and 64.6% on MTEB, priced at $0.00013 per 1k tokens.

The real engineering highlight here is dynamic dimension reduction. Usually, vector stores demand fixed dimensions, forcing developers to balance accuracy against hefty RAM and storage costs.

With these new models, passing a dimensions parameter lets you trim values from the sequence without breaking its ability to represent concepts. In fact, OpenAI notes that 3-large shortened down to just 256 dimensions still beats unshortened 1536-dimension ada-002 on MTEB. If your vector database caps out at 1024, you can now use the large model without headache.

Curing laziness and slashing GPT-3.5 prices

OpenAI also tackled model personalities. The new gpt-4-0125-preview model specifically targets "laziness," where GPT-4 would fail to finish code blocks or leave tasks half-done. It also patches a nagging UTF-8 bug affecting non-English generations.

Meanwhile, GPT-3.5 Turbo is getting its third price cut in a year. The upcoming gpt-3.5-turbo-0125 slashes input costs by 50% down to $0.0005 per 1k tokens, while outputs drop 25% to $0.0015.

To round things out, OpenAI added granular API key permissions (like read-only access) and key-level usage tracking, alongside a new free moderation model called text-moderation-007.

Why it matters

For builders running semantic search and RAG workflows, this update offers an immediate efficiency boost. Slashing embedding costs by 80% while letting you shrink vectors to save database memory makes production-grade retrieval dramatically cheaper. If you are still running ada-002, upgrading is a no-brainer.

Strategically, OpenAI is showing that it understands practical developer bottlenecks better than most. By pairing aggressive pricing with flexible architectural tricks like dimension pruning, it is squeezing both competing proprietary APIs and self-hosted open-source alternatives.

We think the dynamic dimension trick is the smartest move here. Instead of forcing developers into a high-cost storage tier to get top-tier retrieval accuracy, OpenAI removed the infrastructure friction entirely, keeping developers locked firmly into its platform.