Model

codex

6 entries · June 2021 – June 2026

Codex is a name OpenAI has used twice. The first, published in 2021, was a version of GPT-3 fine-tuned on public GitHub code that could turn a plain-English description into working code; it powered GitHub Copilot and introduced the HumanEval benchmark that later code models were measured against. In May 2025 OpenAI reused the name for a different tool — a cloud-based coding agent built on a fine-tuned o3 that works largely unsupervised inside a sandboxed copy of a repository and proposes finished changes as pull requests for review.

Appears alongside

Featured in threads

Tracks

  • Models & capabilities 6
  • Money & business 2
  • Benchmarks & progress 1
  • Culture & impact 1

  1. OpenAI to acquire Ona

    Ona, formerly Gitpod, gives Codex persistent cloud sandboxes so agents can keep working for hours or days after a developer closes their laptop; terms were undisclosed.

    Money & business · Models & capabilities

  2. OpenAI launches Codex, a cloud-based coding agent

    Built on a fine-tuned o3, each task runs in its own preloaded cloud sandbox and proposes a pull request, letting several jobs run at once without a developer at the keyboard.

    Models & capabilities

  3. GitHub Copilot goes on sale

    Priced at $10 a month or $100 a year, it moved from a year-long technical preview used by over 1.2 million developers to a paid product free for students and open-source maintainers.

    Money & business · Models & capabilities

  4. OpenAI publishes Codex, an LLM trained on code powering GitHub Copilot

    A GPT-3 descendant fine-tuned on public GitHub code solved 28.8% of a new benchmark's problems on the first try, rising to 70.2% with repeated sampling.

    Models & capabilities

  5. OpenAI publishes Codex and the HumanEval benchmark

    Codex solved 28.8% of HumanEval's Python problems on a single attempt and 70.2% when allowed 100 samples per problem, against 0% for base GPT-3.

    Models & capabilities · Benchmarks & progress

  6. GitHub launches Copilot in technical preview

    An OpenAI model trained on public code began suggesting whole functions inside the editor — the first mass-market use of a large language model.

    Models & capabilities · Culture & impact