OpenAI’s new Astra model

From a market perspective, the announcement of OpenAI’s Astra model appears to be a technical milestone – one that has yet to be fully validated – but, above all, a significant governance challenge. What is worrying for society is not only the scale of its capabilities, but the intrinsic opacity of its inference processes, which in this new version becomes even more opaque. When a model’s traceability is limited, data-driven decision-making loses a fundamental pillar: the ability to audit how the system weights variables and validates its intermediate hypotheses.

This is made all the more concerning in light of the incident that occurred this summer, which should serve as a warning regarding these organisations’ actual capabilities in terms of the security and reliability of these systems.

This lack of explainability has direct implications for AI security. Without a clear understanding of the model’s internal architecture and reasoning, it is extremely difficult to ensure the mitigation of bias, prevent complex hallucinations, or protect the system against vulnerabilities and adversarial attacks.

The widespread adoption of systems with this level of ‘black box’ complexity demands a paradigm shift; the launch of Astra must urgently drive the introduction of more rigorous regulatory frameworks and risk audits to protect both organisations and the public.

The launch of GPT-6 Astra has taken the AI community by surprise because it represents a significant leap forward compared to its competitors, clearly positioning it as the model that performs best in controlled evaluations. At the same time, it presents society and businesses with a challenge regarding the governance of these models, given that, although the company insists it has made significant efforts to improve alignment, the new techniques being applied may limit the explainability of the model’s behaviour. Added to this is the fact that some of its most sensitive capabilities are beginning to be subject to restricted access and accreditation, which may clash with European sovereignty. If we add to this cybersecurity incidents, such as the one recently suffered by Hugging Face, we can see that the governance of this type of model leaves room for improvement and is far from straightforward.

Statements by OpenAI executives suggesting that Astra marks the dawn of the era of artificial general intelligence (AGI) are a major marketing strategy designed to attract venture capital investment rather than representing the realisation of general intelligence.

These programmes are designed to mimic human language so effectively that they lead the public to project onto them a thinking mind, cognitive abilities or intentions that they do not actually possess. Regardless of how advanced GPT-6 Astra’s computational capabilities or agentic functionalities may be, the model remains a synthetic media generation machine that operates exclusively on linguistic form — the structure and mathematical probability of the data — without any real connection to meaning, let alone a thinking mind, that is to say, an artificial general intelligence comparable to or superior to human intelligence.

In fact, the very concept of ‘general artificial intelligence’ is problematic, as there are a multitude of different definitions. OpenAI has historically defined it in functional terms: highly autonomous systems capable of outperforming humans in most economically valuable tasks. But demonstrating that a model satisfies such a broad definition requires more than simply achieving exceptional results on certain benchmarks or performing spectacular demonstrations.

In fact, independent analyses have pointed out that Astra achieves very mixed results on other benchmarks. For example, competing models such as Anthropic’s Claude 5.1 continue to lead on certain metrics that measure the ability to develop software, whilst Astra actually showed a decline in quality compared to its predecessor, GPT-5.6 Sol, when tackling other benchmarks.

Furthermore, some of its most spectacular results require additional tools, memory and components based on techniques distinct from those of large language models. In other words, it is a neurosymbolic AI rather than a pure LLM. Therefore, we are by no means dealing with general artificial intelligence. Talking about general artificial intelligence, as OpenAI does, is a marketing strategy.

But there is a much more specific problem on which all experts seem to agree: the more capable these systems are of acting autonomously, the more difficult it can be to determine what they are doing. This aspect is particularly relevant in the case of Astra. OpenAI itself acknowledges that the new model presents a reduced ability to monitor its operation, and this is not good news. Security teams, including those at Redwood Research and the UK AI Security Institute (AISI), have highlighted this as a dangerous step backwards in terms of the transparency and accountability of these systems when things go wrong.

What I’m seeing with Astra strikes me as truly impressive. Even working in the field of artificial intelligence, the progress being made in such a short space of time is astonishing. Personally, I think we need to set aside the debate over whether this constitutes general artificial intelligence or not. I believe it has a strong commercial element at the moment. That said, we mustn’t fail to recognise the importance of these advances.

I think we should highlight this new model’s ability to use a computer: it can run programmes, browse the web and complete multi-step tasks. Impressive examples of video game creation and 3D modelling are rapidly emerging. Obviously, these are demonstrations and do not guarantee that it will work the same way in every situation, but my impression is that the transformation in the world of software development is gathering pace.

The results in ARC-AGI-3 are also highly significant. These are small video games unknown to the system, where it must discover the rules to solve them. Astra achieves a virtually perfect score (99.9 per cent) when using a specific configuration that preserves its reasoning and manages the context. With a standard configuration, it achieves 62.7 per cent, more than double the 30.2 per cent published for Claude Opus 5. This does not mean that AGI has been achieved, but it does demonstrate significant progress in complex reasoning and in the systems’ ability and efficiency to learn new tasks on their own.

These results once again highlight the saturation of certain benchmarks and the need to develop more demanding tests. Furthermore, the difference between configurations demonstrates that how the complete system is built matters: the model, the tools and context management. I believe we continue to talk about models, but in reality we are increasingly working with complex systems that combine these elements, and where advances are not only scientific but also engineering-based to ensure this orchestration works.

The new capabilities of this model (and those yet to come) mean we must remain vigilant, particularly in the field of cybersecurity. They can help to fix vulnerabilities, but they can also facilitate attacks. Public authorities and companies providing essential services must prepare for this change. Automation can make it feasible to find and exploit vulnerabilities that previously required a great deal of time and specialist knowledge. At European level, it is important that we address these risks without being left behind in the development or use of this technology. Access to the best systems and the know-how to make the most of them can make a significant difference.