OpenAI’s next major model, GPT-5, is reportedly proving harder to bring over the finish line than expected. According to reporting cited by TechCrunch, the model can perform better than earlier systems, but the improvement has not yet been large enough to justify the enormous costs involved in running it.
GPT-5 is reportedly behind schedule
The development effort for GPT-5 has been running for 18 months under the code name Orion. The Wall Street Journal reported that the work is behind schedule, and TechCrunch noted that this aligns with an earlier report in The Information suggesting OpenAI has been looking for new strategies.
The core issue is not that GPT-5 has failed to improve. The report says it can outperform its predecessors. The problem is that the scale of the improvement has not yet matched the scale of the expense.
That distinction matters. A new flagship AI model is expected to represent a meaningful step forward, especially when it requires extensive training, large amounts of data, and significant resources to operate. If the gains are incremental while the costs are enormous, the release decision becomes more complicated.
Training runs show the cost problem
OpenAI has reportedly completed at least two large training runs for GPT-5. These runs are designed to improve a model by exposing it to enormous quantities of data.
The first reported training run moved more slowly than expected. That was an important signal because a larger run would likely take more time and cost more money. In this context, speed is not just a technical concern; it also affects whether the model can be developed and operated in a way that makes sense.
The reporting describes a familiar tension in frontier AI development: bigger systems may produce better results, but each step up can demand more resources. For GPT-5, the question appears to be whether the extra performance is enough to justify the extra cost.
- GPT-5 has reportedly improved over earlier models.
- The improvement has not yet been enough to justify the cost of keeping it running.
- At least two large training runs have been completed.
- An early run was slower than expected, pointing to higher time and cost demands for larger runs.
OpenAI is broadening its data strategy
The Wall Street Journal also reported that OpenAI is not relying only on publicly available data and licensing deals. The company has reportedly hired people to create fresh data by writing code and solving math problems.
That detail points to one reason GPT-5 may be difficult to advance. When existing data sources are not enough, a company may need new, more targeted material to train or evaluate a model. In this case, the reported examples are code and math problems, both areas where clear answers and structured reasoning can matter.
OpenAI is also reportedly using synthetic data created by another of its models, o1. Synthetic data can expand the material available for training, but the source article does not say how much of it is being used or how central it is to the GPT-5 effort.
The important point is that OpenAI appears to be combining several data sources: public material, licensing deals, human-created work, and model-generated data. That mix suggests the company is searching for ways to produce stronger results without relying on a single pipeline.
Orion may not be the leap many expected
TechCrunch noted that the Wall Street Journal report echoes an earlier report in The Information, which suggested GPT-5 might not deliver as large a leap as previous models. That does not mean the model lacks progress. It means expectations for a next-generation release may be difficult to meet.
For OpenAI, the stakes are tied to both performance and economics. A model can be technically stronger and still fall short of the threshold needed for a major release if it is too expensive to operate relative to its benefits.
The company previously said it would not be releasing a model code-named Orion this year. OpenAI did not immediately respond to TechCrunch’s request for comment on the Wall Street Journal report.
For now, GPT-5 appears to be a case study in how hard it is to keep pushing major AI systems forward. The reported challenge is not simply building a better model. It is building one whose improvement is large enough to make the time, data, training effort, and operating cost worthwhile.