Milestones
- GPT-3 (175B parameters) paper and APIMay 2020Complete.
- ChatGPT launched30 Nov 2022Complete.
- GPT-4 released14 Mar 2023Complete.
- GPT-4o native multimodal model13 May 2024Complete.
- o1-preview reasoning model12 Sep 2024Complete.
- GPT-5 released7 Aug 2025Complete.
- GPT-5.5 released23 Apr 2026Complete.
- GPT-5.6 family (Luna, Terra, Sol) released9 Jul 2026Complete.
- AI research intern-level system (stated goal)Target Sep 2026Current milestone.
- Fully automated AI researcher (stated goal)Target Mar 2028Not yet reached.
Most important updates
- 9 Jul 2026
- 23 Apr 2026
- 7 Aug 2025
- 12 Sep 2024
- 14 Mar 2023
Upcoming
- Sep 2026AI research intern-level system (stated goal) (next)
- Q4 2026First AMD MI450 (Helios) capacity expected online for OpenAI
- Mar 2028Fully automated AI researcher (stated goal)
Current obstacles
- Reliability on long tasksSmall mistakes add up over hours-long tasks. Getting models to notice and fix their own errors is unsolved.
- Running out of good dataHuman-written text online is finite. Progress now leans on AI-made data, practice tasks and paid experts.
Physics limits
- Returns shrink as a power lawError falls only as a small power of training compute: each fixed step of improvement needs roughly 10x more compute, energy and money than the one before.
- Trained for likely text, not true textThe model learns to produce plausible words. Facts seen only once or twice in training can't be recalled reliably, so some rate of confident errors is built into the method.
- Only checkable skills train wellReasoning is learned by rewarding answers a program can grade, such as maths and code. Where no automatic checker exists (strategy, taste, open research) the signal is weak.
How it works
Pretrain
Guess the next word
Trained to predict the next word across trillions of words of text and code. Doing that well forces it to learn facts, grammar and patterns of reasoning.
Align
Learning to be helpful
People rank sample answers, a reward model learns their preferences, and the main model is tuned to score well while following a written behaviour spec.
Reason
Thinking before answering
Reasoning versions work through a hidden chain of thought. They practise on maths and code, where answers can be checked.
Route
Picking the right model
A router sends each request to a fast model or a deep-thinking one, trading speed for accuracy.
Papers & demos
- Sep 2024blogLearning to Reason with LLMsIntroduced o1: answers got better with more training and more thinking time.
- Mar 2023paperGPT-4 Technical ReportShowed a big multimodal model's progress could be predicted, and that it passed hard exams.
- Mar 2022paperTraining language models to follow instructions with human feedback (InstructGPT)Set the recipe for turning a raw model into a helpful assistant using human ratings.
Update log
Thu 9 Jul
- Major: BlogSoftware
Thu 23 Apr
- Major: BlogSoftware
Thu 7 Aug 2025
- Major: BlogSoftware
Thu 12 Sep 2024
- Major: BlogSoftware
Tue 14 Mar 2023
- Major: BlogSoftware
Wed 30 Nov 2022
- Major: BlogSoftware
About OpenAI
OpenAI
ChatGPT, GPT models, Codex, Stargate
OpenAI makes ChatGPT and the GPT models behind it. ChatGPT's launch in November 2022 made AI chatbots mainstream. It also builds huge data centers through the Stargate project.
- Makes ChatGPT and the GPT-5.x models. GPT-5.6 came out in July 2026.
- Raised $122B at an $852B valuation in March 2026. Amazon, NVIDIA and SoftBank were the biggest investors.
- Lining up tens of gigawatts of computing power through Stargate and chip deals with NVIDIA, AMD and others.
- Founded
- 201511 yrs
- Headquarters
- United States
- Status
- Private
- Valuation
- $852BprivateMar 2026
- Raised
- $179B4 rounds
- Last round
- Venture round · $122BMar 2026$852B post
- Coverage
- 4 programs · 24 updateslatest 9 Jul 2026checked 25 Sep
- Lead investors
- AmazonNVIDIASoftBankThrive CapitalMicrosoft
- Partners
- Broadcom
- Investors
- MicrosoftSoftBankAmazon
- People
- Sam AltmanCEO and co-founderGreg BrockmanPresident and co-founderJakub PachockiChief Scientist