"GPT-4 was already finished before GPT-3.5 was released. That explains a lot about OpenAI." The X post by Adam Holter (@AdamHoltererer), which drew 110,000 views and 1,000 likes, is a single short sentence, but it makes you reread OpenAI's whole release strategy. Is it true? We cross-checked it against official sources.
Seven Months After Completion, Before the World Saw It
Start with the key fact. OpenAI's GPT-4 system card, published in March 2023, has a decisive sentence: "Since it finished training in August of 2022, we have been evaluating, adversarially testing, and iteratively improving the model and the system-level mitigations around it." So training finished in August 2022, and the public release came about seven months later, on March 14, 2023.
CEO Sam Altman also said in April 2023 that "we spent more than six months after finishing training GPT-4 before releasing it," explaining the gap as safety and alignment research. (Business Insider) Bill Gates's recollection lines up too. In his March 2023 post "The Age of AI Has Begun," Gates wrote that in summer 2022 he watched the OpenAI team have GPT-4 answer 59 of 60 AP Biology exam questions correctly.
| When | Event | Source |
|---|---|---|
| March 2022 | code-davinci-002, the base model of the GPT-3.5 series, released first through the API | Hacker News · VentureBeat |
| August 2022 | GPT-4 finishes training (stated in the system card) | OpenAI GPT-4 |
| August–September 2022 | GPT-4 demonstrated to Bill Gates (59 of 60 AP Biology questions correct) | Gates Notes |
| November 30, 2022 | ChatGPT (GPT-3.5-based) released — 1 million users in 5 days | OpenAI · Wikipedia |
| March 14, 2023 | GPT-4 released — a surprise, unannounced launch | OpenAI · The Verge |
Before GPT-3.5? The Naming Trap
Read literally, Holter's claim means "GPT-4 (finished August 2022) was completed before GPT-3.5 (released November 2022)." ChatGPT launched on November 30, 2022, on a GPT-3.5-series model, so the comparison holds. (OpenAI)
But there's a catch. code-davinci-002, the base model of the GPT-3.5 series, was already available to developers through the API in March 2022. (Hacker News) The name "GPT-3.5" was attached in November 2022 with text-davinci-003 and ChatGPT, but models of that lineage had been out in the world five months before GPT-4 finished training. So it's true for "GPT-3.5 as a product" but not accurate for "GPT-3.5 as a model family." The gap between name and substance is itself a clue to understanding OpenAI.
What the Seven-Month Hold Left in the Market
This timeline suggests three practical things.
First, OpenAI ships "the model that best fits the market" before "the best model." Even with a stronger GPT-4 already in hand, it threw out ChatGPT (GPT-3.5) first and generated an explosive response. The result was 1 million users in five days. (Wikipedia) It means that from a developer's perspective, the latest model can be held back at any time, and the release order can differ from the performance order.
Second, names are strategy. GPT-3.5 reads less as a standalone product than as "laying the groundwork" for the GPT-4 that would arrive four months later. GPT-4 came out as an unannounced surprise, (The Verge) and OpenAI didn't disclose the model size, training data, or compute scale. MIT Technology Review interpreted that as "a shift from a nonprofit research lab to a for-profit tech company." (MIT Technology Review)
Third, safety alone may not explain the seven-month gap. OpenAI stressed "more than six months of safety and alignment work," but some commentators see that period as closer to post-processing for the product (fine-tuning, RLHF) and internal validation. (GreaterWrong) The line between "safety" and "commercialization prep" is hard to see from outside. That opacity is itself a defining trait of OpenAI's release strategy.
Limits of Verification and Open Questions
Confirmed facts: GPT-4 finished training in August 2022 (OpenAI system card), ChatGPT was released on November 30, 2022 (OpenAI, Wikipedia), and GPT-4 was released on March 14, 2023 (OpenAI, The Verge). All three dates were cross-checked against official sources.
Partially confirmed: The 110,000 views and 1,000 likes on Holter's tweet couldn't be re-verified directly on the original page because of X's public access restrictions, so we relied on the figures we were given.
Not confirmed: (1) The body of the GPT-4 technical report (arXiv 2303.08774) doesn't state the training completion date. "August 2022" rests on the system card. (2) We couldn't find Greg Brockman's "ChatGPT fourth anniversary retrospective" post as of August 2026 (the fourth anniversary is November 2026). However, according to foreign press reports, Brockman said around August 9, 2026 that "GPT-4 finished training four years ago today." (foreign press report) (3) Holter's interpretation that "this explains a lot about OpenAI" is speculation. Whether the pattern of releasing named products in sequence while a finished model sat waiting was intentional hasn't been confirmed.