OpenAI is moving faster with new language models, and that speed is putting its safety process under sharper scrutiny. According to a Financial Times report, GPT-4 underwent six months of testing, while testers now have just days to evaluate the new "o3" model.
The concern is not only that the timeline has changed. It is that the change is happening as the models become more powerful and as possible misuse, including biological or security-related uses, becomes a more serious part of the risk discussion.
Why the testing window matters
Safety testing is the part of AI development where a company looks for ways a model could be misused or behave dangerously before it is released more widely. In OpenAI's case, the source article says people involved in the process reported less thorough testing and insufficient resources.
That matters because the testing burden grows with the model's capability. A weaker model may be easier to assess because it can do less. A more capable model can raise harder questions, especially when the risks involve specialized areas such as biological or security-related misuse.
The contrast is stark: GPT-4 had six months of testing, while the new "o3" model is being evaluated in days. The source does not say that a shorter window automatically means the model is unsafe. But it does show why outside observers, testers, and regulators may ask whether speed is overtaking depth.
The competitive pressure behind faster releases
The shorter testing period comes as OpenAI faces pressure to keep up with companies including Meta, Google, and xAI. The source says OpenAI wants to accelerate releases to maintain that pace.
This creates a familiar tension for companies building advanced AI systems. Faster launches can help a company stay competitive, but safety work often requires time, experts, and repeated evaluation. When the release cycle compresses, the hard question becomes what gets reduced, automated, delayed, or left uncertain.
In this case, the Financial Times report points to people involved in the process saying the testing was less thorough and under-resourced. OpenAI, however, points to efficiency gains from automated testing procedures.
Specialized risk tests are resource-intensive
OpenAI previously committed to specialized tests designed to check for potential misuse, including developing biological weapons. The source article describes these procedures as resource-heavy because they require custom datasets, fine-tuning, and external experts.
That kind of testing is different from a simple checklist. It can require creating the right evaluation material, adapting a model for the test, and bringing in people with relevant expertise. When timelines shrink to days, those requirements become harder to fit into the process.
The Financial Times report says this testing was only performed on older, less capable models. It remains unclear how newer models such as o1 or o3-mini would perform under similar conditions.
The o3-mini safety report also leaves a gap. According to the source, OpenAI mentions that GPT-4o could solve a specific biological task after fine-tuning, but it does not provide results for newer models.
The debate over testing "checkpoints"
Another issue is the use of testing "checkpoints," which are intermediate versions of models that are still being developed. A former technical employee described this as bad practice, while OpenAI says the checkpoints are nearly identical to the final models.
The disagreement matters because the model being tested should represent the model that users will actually encounter. If a checkpoint changes meaningfully before release, earlier testing may not fully describe the final system. If OpenAI is correct that the checkpoints are nearly identical, the practical gap may be smaller.
The source does not resolve that dispute. It shows that the practice itself has become part of the broader question around whether AI safety testing is keeping pace with model development.
Regulation is starting to catch up
For now, the source says there are no mandatory global rules for AI safety testing. Companies such as OpenAI have made voluntary commitments to authorities in the US and UK, but those commitments are not the same as a global testing requirement.
That is expected to change when European AI regulations take effect later this year. Those rules will require providers to formally evaluate their most powerful models for risks.
Until then, much depends on company processes, public documentation, and trust in internal judgment. Johannes Heidecke, who leads OpenAI's safety systems, says OpenAI has found a good balance between speed and thoroughness. The company also says that although there is no standardized requirement for processes such as fine-tuning, it follows best practices and documents them transparently.
The central issue is now clear. OpenAI says automation and better procedures can support faster development. Critics and people involved in the process say compressed testing and limited resources can weaken the safety review. As models become more capable, the distance between those two views will become harder to ignore.