Drafted with Claude from official web pages (there is no paper) on 15 October 2026; reviewed by Zeeshan Arif. Read from openai.com (OpenAI blog post of 30 November 2022, read from the Web Archive copy of that day). Also read are OpenAI's ChatGPT API announcement of 1 March 2023 (Web Archive copy) and its current API deprecations page. OpenAI's model index for researchers, which described each GPT-3.5 model, is no longer online.
What it changed
GPT-3.5 is the name OpenAI gave to a series of models that followed GPT-3. OpenAI offered them as products and published no research paper about them. ChatGPT was fine-tuned from one of them, and the cheaper gpt-3.5-turbo API model made the same model available to developers.
The problem before it
GPT-3 was a base model: it continued text, and users had to coax it with careful prompts. InstructGPT had shown that fine-tuning with human feedback made a model follow instructions far better. But that was a research result. A model that followed instructions, held a conversation, and was cheap enough to run for millions of people still had to be built and served.
There is no paper for GPT-3.5. OpenAI described the series in a page called the “model index for researchers”, which is no longer online, so this summary uses only the official pages listed under Links.
The idea
The mechanism. OpenAI’s pages describe GPT-3.5 only in outline:
- A series of models. OpenAI’s ChatGPT announcement says ChatGPT “is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022”. It does not say which model, or how large.
- Trained the InstructGPT way. ChatGPT was trained with reinforcement learning from human feedback, “using the same methods as InstructGPT, but with slight differences in the data collection setup”:
- Human trainers wrote example conversations, playing both the user and the assistant.
- Trainers ranked several alternative replies from the model, and a reward model learned from these rankings.
- The model was then fine-tuned against the reward model with the PPO algorithm. OpenAI repeated this several times.
- The hardware. OpenAI says ChatGPT and GPT-3.5 were trained on an Azure AI supercomputing infrastructure.
- Served through an API. In March 2023, OpenAI released gpt-3.5-turbo, which it said “is the same model used in the ChatGPT product”. Instead of plain text, it reads a list of chat messages.
The notation. None. The sources read here give no size, architecture or training data for the GPT-3.5 models.
What it showed
Reported by the sources. These are OpenAI’s own statements, not peer-reviewed results.
- Cheaper to run. OpenAI said it cut the cost of ChatGPT by 90% between December 2022 and March 2023. It priced gpt-3.5-turbo at $0.002 per 1,000 tokens, “10x cheaper than our existing GPT-3.5 models”.
- Useful beyond chat. OpenAI said early testers moved from text-davinci-003 to gpt-3.5-turbo “with only a small amount of adjustment” to their prompts.
- Better answers to a false premise. The ChatGPT announcement compares ChatGPT with InstructGPT (text-davinci-002) on the request “Tell me about when Christopher Columbus came to the US in 2015”. ChatGPT pointed out that Columbus died in 1506, while InstructGPT answered as if the visit had happened.
- Used in research. The chain-of-thought paper tested text-davinci-002 and code-davinci-002. It says the text-davinci models “presumably correspond to” InstructGPT models. With chain-of-thought prompting, code-davinci-002 solved 63.1% of GSM8K maths problems.
Limits, and what came later
The limits come from OpenAI’s ChatGPT announcement and from the gaps in the sources:
- Plausible but wrong answers. OpenAI says ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. It gives three reasons this is hard to fix, including that reinforcement learning has no source of truth.
- Sensitive to wording. A slight rephrase of a question can change whether it answers.
- Too long. It is often verbose and overuses some phrases. OpenAI says trainers prefer longer answers that look more complete.
- Guesses instead of asking. It usually guesses what an unclear question means, instead of asking.
- Not fully safe. It sometimes responds to harmful instructions or shows bias.
- Little is published. The sources read here give no model sizes, data or training details, and the page that listed the GPT-3.5 models has been removed.
- Inconsistent records. OpenAI’s deprecations page lists text-davinci-002 and text-davinci-003 under “InstructGPT models”. It gives two shutdown dates for code-davinci-002: 23 March 2023 in one section and 4 January 2024 in another.
- Retirement. text-davinci-002, text-davinci-003 and code-davinci-002 were retired on 4 January 2024. The page says the remaining gpt-3.5-turbo model shuts down on 23 October 2026.
ChatGPT was fine-tuned from a GPT-3.5 model. GPT-4 succeeded the series. Alpaca trained an open model on data generated with text-davinci-003, and phi-1 used synthetic textbook data generated by GPT-3.5.
What to know first
- GPT-3, the base model.
- InstructGPT, the training method behind the series.
- What an API is: a way for other programs to send text to a model and get its reply.
How to read it
There is no paper to read. Instead:
- Read the “Methods” and “Limitations” sections of OpenAI’s ChatGPT announcement.
- Read the “ChatGPT API” part of the March 2023 announcement.
- For the model names and dates, look at the 2023 and 2026 entries on the deprecations page.
- For the method itself, read the InstructGPT summary.
Expect about 30 minutes for the steps above (an estimate).
Links