OpenAI Releases text-embedding-3 Models and Embeddings API Enhancements

New Model Architecture and Capability Upgrades

OpenAI officially released the text-embedding-3 family, replacing text-embedding-ada-002 and the earlier text-embedding-3 preview. It comprises two variants: text-embedding-3-small (default dimension 512, max 2048) and text-embedding-3-large (default dimension 1024, max 3072), both supporting user-specified output dimensions (multiples of 8, range 8–3072) at inference time for precise trade-offs between accuracy and cost.

Key Performance Metrics

• Achieves a composite MTEB (Massive Text Embedding Benchmark) v1.1 score of 69.4 on text-embedding-3-large — notably higher than text-embedding-ada-002’s 62.1 — across Retrieval, Clustering, and PairClassification tasks;
• Demonstrates empirically improved support for 100+ languages including Chinese, Japanese, and Arabic, with average cosine similarity gains of 12.7% on cross-lingual retrieval tasks (e.g., XNLI-based retrieval);
• Supports up to 8191-token context length (aligned with GPT-4 Turbo), enabling robust embedding of longer document chunks.

Critical API-Level Updates

• Introduces a new batch endpoint (/v1/embeddings/batch) supporting up to 2048 texts per request, delivering 3.2× higher throughput;
• Implements dimension-aware token pricing: billing is based on actual output dimension (e.g., 512-dim output billed as 512 tokens, not fixed 1536);
• Adds usage object in responses, including prompt_tokens, total_tokens, and model, enabling granular monitoring and cost attribution;
• Supports input_type parameter ("query" or "document") to optimize vector space alignment for RAG applications.