gpt-4-0314

gpt-4-0314 is the launch snapshot of GPT-4, released by OpenAI on Mar 14, 2023 — the system card’s “GPT-4-launch,” i.e. GPT-4-early plus the safety-focused RLHF pass. It shipped to ChatGPT Plus subscribers and a waitlisted API on day one; Microsoft simultaneously confirmed Bing Chat had been running a GPT-4 lineage model since February.¹

Launch claims centered on professional-exam performance — most famously a simulated bar exam score “around the top 10% of test takers” against GPT-3.5’s bottom 10%.¹ (Eric Martínez’s later re-analysis put it “below the 69th percentile” against July test-takers, and around the 48th on the essay portion — the top-10% framing held only against February repeat-takers.²) Context length was 8,192 tokens, with a gpt-4-32k variant. GPT-4 was announced as multimodal, but at launch “image inputs [were] still a research preview”¹ and stayed so until GPT-4V arrived in ChatGPT in late September 2023³ — as deployed, 0314 was text-only.

The launch also marked a turn in AI discourse: OpenAI published no architecture, parameter-count, or training-data details, citing “the competitive landscape and the safety implications of large-scale models,”⁴ and within two weeks the Future of Life Institute’s open letter called for a pause on training systems “more powerful than GPT-4.”⁵

Lifecycle

• Superseded in the API by gpt-4-0613 on Jun 13, 2023; gpt-4 requests pointed at the newer snapshot thereafter.
• The ChatGPT product clock ran separately: “GPT-4” left ChatGPT on Apr 30, 2025, fully replaced by GPT-4o (by then it was serving 0613-lineage weights),⁶ while 0314 stayed reachable in the API until its retirement in March 2026.

Further reading

• “On GPT-4” — Zvi Mowshowitz (Mar 15, 2023) — the launch-week roundup
• Chatgpt-4 — Cyborgism Wiki — on the assistant persona’s mode collapse (“As an AI language model…”), as against gpt-4-base and Sydney