Transformer-based Encoder-Decoder Models

Production deployments have converged almost entirely on decoder-only architectures — GPT, Llama, Mistral. Encoder-decoder models like T5 and BART are widely treated as legacy, despite holding seq2seq benchmark records through 2023.

For bounded generation — document summarization, translation, structured extraction — encoder-decoder models outperform decoder-only at equal parameter counts. FLAN-T5-large handles summarization pipelines at roughly 10x lower inference cost than GPT-4o. Defaulting to a chat-optimized model for every generation task is architectural laziness dressed up as pragmatism.

Teams running high-volume, fixed-schema summarization or translation workflows should benchmark a fine-tuned T5 variant before renewing API spend. If you're doing open-ended generation or multi-turn dialogue, skip this entirely.