AI Radar

GPT-3

Retired · Language · Milestone

GPT-3 is an autoregressive language model from OpenAI with 175 billion parameters, introduced in May 2020 in a paper that applies it to tasks given as text instructions or a few examples in the prompt. OpenAI offered it only through its commercial API, launched in private beta in June 2020, and shut down the base GPT-3 API models in January 2024. [1] [2] [3] [4] Primary source

Timeline of GPT-3 →

Claims and evidence

  • Announced Primary source[2]page date [1]arXiv submission history, v1
  • Released API launched in private beta; access required approval. Primary source[3]page date; models with weights from the GPT-3 family
  • Deprecated Primary source[4]2023-07-06: GPT and embeddings
  • Retired Primary source[4]2023-07-06: GPT and embeddings, Base GPT models
  • Status RetiredThe base GPT-3 API models ada, babbage, curie and davinci were shut down on 2024-01-04. Primary source[4]2023-07-06: GPT and embeddings, Base GPT models
  • Successor of GPT-2The paper says GPT-3 uses the same model and architecture as GPT-2, except for alternating dense and locally banded sparse attention, at a much larger scale. Primary source[1]section 2.1 Model and Architectures
  • Change · Size 175 billion parameters, against 1.5 billion for the biggest GPT-2 version.Compared with GPT-2 Primary source[1] [5]
  • Change · Context length Context window doubled to 2048 tokens, from 1024 for GPT-2.Compared with GPT-2 Primary source[1] [6]
  • Change · Training data Training data adds filtered Common Crawl, two internet-based book corpora and English Wikipedia to an expanded WebText.Compared with GPT-2 Primary source[7]
  • Change · Availability Offered only through the OpenAI API, with access by approval, instead of as downloadable weights.Compared with GPT-2 Primary source[3]
  • Input text Primary source[3]text in, text out interface
  • Output text Primary source[3]text in, text out interface
  • Open weights No Primary source[3]FAQ: Why did OpenAI choose to release an API instead of open-sourcing the models? [7]Model Use
  • Context window 2048 tokens tokens Primary source[1]section 2.1
  • Parameters 175 billionThe paper trained eight sizes from 125 million to 175 billion parameters and calls only the largest GPT-3; OpenAI's model index maps GPT-3 175B, 6.7B and 1B to the API models davinci, curie and babbage. Primary source[1]abstract [7]Model Details
  • Access API Primary source[3] [7]Model Use

Lineage

Predecessors

  • GPT-2 · 14 February 2019

Successors

Based on

Not derived from another model.

Variants and derived

  • InstructGPT · 27 January 2022 · derived (fine tune)

Siblings

None recorded.

All ancestors

  • GPT · 11 June 2018
  • GPT-2 · 14 February 2019

All descendants

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from OpenAI on AI Radar →