Skip to content

Prism

external · status: active · focus: drafting · discipline: general · started: 2026

Project page: https://openai.com/prism/

Source: projects/landscape/prism.yml

Positioning

OpenAI's free AI-native LaTeX workspace for scientific writing and collaboration, launched late January 2026 on Crixet, a cloud LaTeX platform OpenAI acquired. GPT-5.2 (upgraded to GPT-5.3 with Codex CLI in March 2026) operates inside the document with access to the paper's structure, equations, citations, and figures, covering drafting, revision, real-time collaboration, arXiv-backed referencing, and publication prep. Sits in the writing layer of the RISE pipeline — the manuscript-workspace counterpart to workbench products like Claude Science.

Distinctive contribution

First big-lab product to make the manuscript itself the AI-native surface: the model is embedded in the LaTeX project rather than bolted on as a chat sidecar, and the workspace is free with unlimited projects and collaborators on a personal ChatGPT account. Converts whiteboard/handwritten sketches to LaTeX, auto-finds and formats arXiv references, and — since the March 2026 Codex integration — runs Python/R in-workspace with compile-check-revise self-repair loops for LaTeX errors.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Six stages centered on the writing layer (reference search, in-workspace analysis/code since March 2026, drafting, revision, publication prep); nothing upstream of the manuscript.
Autonomy level 1 Copilot: the researcher authors the paper; the in-document model drafts, edits, and self-repairs LaTeX on request, with the human approving changes.
Architectural transparency 1 Model identity (GPT-5.2/5.3-Codex), Crixet foundation, and in-document context integration disclosed; no prompts, code, or architecture documentation beyond marketing.
Inputs supported 2 Multiple input forms (LaTeX projects, natural-language and voice instructions, sketch photos, data files) plus arXiv literature access; no private-corpus integration.
Outputs / reproducibility 2 Cloud projects persist LaTeX source, bibliographies, and (post-March) executed code and generated figures; no Git integration, limited version-history transparency, no reproducibility manifest.
Internal evaluation 0 No published evaluation, benchmark, or systematic quality evidence in launch or upgrade coverage at scoring date.
Openness 1 Closed-source, but the hosted service is free with a personal ChatGPT account (unlimited projects/collaborators); advanced AI features expected to move behind paid tiers.
Maturity / traction 2 Production launch (Jan 2026) on an acquired platform, open globally to ChatGPT Free/Go/Plus/Pro with a major March upgrade; no adoption numbers disclosed.
Cross-family policy 0 OpenAI models only (GPT-5.2, then GPT-5.3 with Codex CLI); no cross-family configuration.
Runtime assurance 1 Compile-check-revise self-repair loops and arXiv-grounded citation insertion; no claim-audit, fact-check, or figure-integrity gates.
Cross-platform portability 0 Single hosted web surface tied to one provider; no self-hosting, IDE integrations, or alternative runtimes.

Scored on 2026-07-23. See the evaluation rubric.

Tags

Pipeline stages: literature-discovery data-analysis code-generation paper-drafting revision-editing dissemination

Architectural features: single-llm tool-use iterative-loop

Inputs: latex-project natural-language-instructions whiteboard-sketches data-files

Outputs: latex-manuscripts formatted-bibliographies figures compiled-pdfs

Data sources: user-provided

Knowledge sources: arxiv

Limitations

  • Closed-source hosted service: manuscripts live in OpenAI's cloud with only the standard ChatGPT training opt-out; no self-hosting or audit path.
  • No native Git integration and limited version-history transparency; early reviews describe it as a functional silo disconnected from external codebases.
  • Locked to OpenAI models, and the free tier is strategic — advanced AI features are slated to require paid ChatGPT subscriptions.