Strawberry and Orion: What Was Reported About OpenAI’s Next Models in 2024

In August 2024 I am looking at reports about OpenAI’s projects called Strawberry and Orion. The Information cites a person with direct knowledge who says Strawberry technology was demonstrated to US national-security officials that summer. That is a reported allegation, not public confirmation that a finished GPT-5 was shown or was imminent.

The reporting treats Strawberry as a reasoning-focused project and Orion as a separate future flagship model. I also discuss Q*, an earlier reported research name, and wonder whether an old cryptic post from the anonymous account Jimmy Apples anticipated Orion. Such posts are ambiguous and cannot verify insider knowledge or establish that every codename refers to the same system.

The reported goal of Strawberry is better performance on difficult tasks such as mathematics through additional reasoning time. I describe that in terms of multi-step thinking rather than an immediate answer. The internal mechanism was not publicly established by these articles, and a model’s longer processing should not be assumed to work like human thought.

The Information reports a demonstration to national-security officials and speculates it might reflect an effort to engage policymakers. Its purpose, the technology’s maturity and any government assessment were not disclosed in the cited account. I discuss open-weight releases and argue that proprietary model weights are difficult to protect indefinitely, but that is my view rather than a demonstrated inevitability.

One reported use for Strawberry is generating high-quality synthetic training examples for Orion. I think this matters because developers face competition and constraints in acquiring useful real-world training data. My claim that practically all accessible internet data is exhausted is too broad; the availability and quality of training data depend on source, permissions and selection.

There is also an open question about synthetic data: if examples are generated by existing models, can they improve a successor, or will errors and limitations be recycled? Better examples could help in some settings, but a report about planned use does not establish the quality of the data or prove hallucinations will disappear.

The Information describes an effort to distill Strawberry for a chat product before Orion. I infer that a large reasoning system might be too slow or expensive for some consumer uses, but its precise latency and implementation were not publicly established. Taking more time can help on hard tasks, while changing the cost and responsiveness users experience.

Possible applications include difficult coding problems and agents that work asynchronously, where an immediate answer matters less. I imagine users choosing when to use a slower reasoning mode. These are product ideas, not features the reporting confirmed would ship in a specific form.

A second Information article describes OpenAI pursuing a reasoning product and raising capital. It says employees had shown colleagues examples such as solving the New York Times Connections puzzle. I interpret this as potentially using intermediate reasoning or search, but the exact technique was not confirmed; it would be premature to name a particular algorithm as fact.

The article also reported approximately $283 million in monthly revenue and an $86 billion private valuation at the time, with likely larger monthly losses. Those historical figures are attributed reporting and are not independently verified here. I see competition from open-weight models, Anthropic, xAI and search products increasing pressure on OpenAI.

I argue that models such as Llama 3 are becoming competitive with GPT-4o for many uses, while cheaper models may be run or fine-tuned independently. That is my assessment, not a demonstrated equivalence across tasks: quality, inference costs and hardware needs vary substantially with the model and workload.

The article again attributes the possible use of Strawberry-generated data to a person familiar with the work and suggests it might improve Orion’s reliability. I see hallucinations as a serious barrier in enterprise and high-stakes applications. Retrieval, improved prompts, independent checks and agent workflows can help, but they do not guarantee error-free answers.

A quoted AI startup leader suggests cleaner examples might reduce ambiguity and improve reasoning. I would like a system that solves logic problems reliably on the first try, but training data quality is only one part of the problem, and the report does not show that any model has eliminated hallucinations.

I am skeptical that one private lab can indefinitely maintain a huge lead when research spreads across a broad community. That is my judgment, not a reliable forecast of future model performance; private data, infrastructure and unpublished work can also affect results.

The article points to mathematics and engineering as potentially valuable uses for stronger reasoning. I note Google DeepMind’s 2024 International Mathematical Olympiad result: its AlphaProof and AlphaGeometry 2 systems solved four of six problems at a combined silver-medal standard, using formalized problems and specialized methods ([DeepMind](https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/)). That does not mean a general chatbot independently earned a medal under ordinary contest conditions.

I also mention Anthropic’s improvements in coding and visual question answering and the popularity of coding tools paired with Claude. The landscape was competitive in 2024, but user preferences and performance vary by task. I object to dismissing multi-step prompting as a cheap trick: breaking a problem down can be useful, even when it adds latency or cost.

The Information traces the reported research to earlier work associated with Ilya Sutskever and other OpenAI researchers. The video speculates about internal safety concerns and departures; those reports do not reveal a specific undisclosed breakthrough or prove a particular reason for every departure.

I finish with a New York Magazine article about the excitement around leaks, codenames and online speculation. It mentions one of my own videos, which I find flattering. Speculation can be fun, but anonymous posts, dramatic titles and predictions are not confirmation of what a lab has built.

Later developments give this story a clearer timeline: OpenAI introduced [o1-preview in September 2024](https://openai.com/index/introducing-openai-o1-preview/), and The Information connected it to Strawberry; OpenAI released a [GPT-4.5 research preview in February 2025](https://openai.com/index/gpt-4-5-system-card/). Those releases do not independently verify every earlier report about demonstrations, training data or internal codenames.