Codex and Copilot

OpenAI fine-tuned a language model on public code, and GitHub embedded it in the editor as an assistant completing lines as you type.

Why it matters

The first mass-market product on a large language model appeared in a code editor rather than a chat, some seventeen months before ChatGPT, counted from the Copilot technical preview of June 2021.

Code proved a convenient testbed: it is structured, abundant in the open, and partly checkable by running it. The product immediately raised the questions that later faced the whole industry: who owns the training data, who is liable for an error, what to do about reproduced fragments of someone else's code. Lawsuits began in 2022.

Event record

Event date

August 2021

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

ID

evt-0242

The model was announced in August 2021; GitHub Copilot entered technical preview in June 2021.

Sources

primary OpenAI Codex
OpenAI · Published August 2021
primary Introducing GitHub Copilot: your AI pair programmer
GitHub Blog · Published June 29, 2021

Related events

Earlier

Related HumanEval: code checked by tests 2021links to this record
The preprint that introduced HumanEval describes Codex and says a distinct production version of it powers GitHub Copilot.
- Evaluating Large Language Models Trained on Code (Chen et al., arXiv:2107.03374v1) · arXiv (OpenAI)
Extends GPT-3 2020
The same model, fine-tuned on public code.
- OpenAI Codex · OpenAI

Later

Builds on The first suit over training on other people's code 2022links to this record
The suit challenges the very product this record describes: a model trained on public code and embedded in the editor.
- Class Action Complaint, J. Doe 1 and J. Doe 2 v. GitHub, Inc., Microsoft Corporation, OpenAI, Inc. et al., Case 3:22-cv-06823 · Joseph Saveri Law Firm and Matthew Butterick
Related ChatGPT 2022
The first mass product on a large language model, a year before ChatGPT.
Related Computer use 2024links to this record
From completing code in an editor to operating the computer itself.
Related ChatGPT as an agent 2025links to this record
From completing a line of code to carrying out multi-step work unattended.
Related Codex-Spark on Cerebras chips 2026links to this record
A shared name, Codex, and a shared field: a model for coding. The records do not establish technical descent between the 2021 model and Codex-Spark.