Altman's 2022 plan: release a GPT-3 model 'before Stability or someone else does'

An October 1, 2022 email from Sam Altman to the OpenAI board, filed this month as an exhibit in Musk v. Altman, is the open-weights containment playbook in plain text. The proposal, sent in the window between Stability AI's first open-weight image models and the launch of any open-weight language model competitive with a frontier lab, called for OpenAI to release a model roughly as capable as 2020's GPT-3 that would run locally on consumer hardware — the same shape as the open-weight models that followed, in which the model file is downloadable rather than closed behind an API — and to do it "before Stability AI, the Stable Diffusion maker, or someone else does."

The strategic intent is in the next two operational lines, which Simon Willison's July 20, 2026 post reproduces from the filed exhibit. Altman wrote that releasing the model would "discourage others from releasing similarly-powerful models" and "makes it harder for new efforts to get funded." Both lines are about the receiving landscape, not the release. The model is the lever; the goal is fewer and weaker competitors. Read the email as a deployment plan, and the desired outcome is not more open-weight models in the world — it is fewer.

The exhibit surfaced through discovery in Musk v. Altman, a federal jury trial in California that has been running since late April 2026, and reached the wider web via a tweet by @TechEmails. The Verge's running exhibit-by-exhibit catalog confirms the email is a filed court exhibit, not a leaked memo or a journalist's reconstruction. The document was put before a jury under oath, with both sides able to challenge its authenticity and context — a different chain of custody than an anonymous leak.

Two years and seven months later, Altman told reporters a different story. In April 2025, PyMNTS reported that OpenAI was weighing a "different open-source strategy" — a public framing that, on its own, sounds like the 2022 email written from a kinder chair. Fortune, also in early April 2025, placed the pivot inside a $300B valuation moment and the immediate aftermath of DeepSeek's release — the first widely deployed open-weight model to land within striking distance of a frontier lab's closed system, in the same week a Ghibli-style image filter was running on top of it. Whatever the board was told in 2022, the company's 2025 posture had to absorb the open-weights wave it had not pre-empted.

The 2022 email is not a verdict on current OpenAI policy, and the trial has not concluded. It is a dated document with named recipients, surfaced under oath, and it tells readers what a frontier lab was privately telling its board about local-capable open-weight releases four years before the open-weights wave actually arrived. The proposal was not executed as written. The receiving landscape it was written to constrain is the one the open-weights community has been building since.