OpenAI launched GPT-6 Astra this week via a gradual roll-out, describing it as the most intelligent and aligned model it has built. Company president Greg Brockman went further and closed the press briefing with "Welcome to the AGI era."
The benchmarks support the view that Astra is a major new entrant. It scored 97.6% on FrontierMath Tier 4, which is remarkable given where this benchmark was just a year ago. Its largest gains come when it must explore an unfamiliar environment, operate software, preserve information and complete a sequence of connected actions. ARC Prize found that Astra developed compact symbolic representations for individual games within the ARC benchmark. In some cases, it created a small domain-specific language to describe objects, states and actions. It then used this representation to plan its next move. The continuous-harness version used fewer actions than the median human tester on 96% of levels, the first model to beat the human baseline.
“GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. ... In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels.”
Astra driving Blender on its own: a full architectural scene, modelled, planted and lit, with a 250 frame flythrough queued on the timeline. The same computer-use ability extends to animations and playable games.
This behaviour has exciting economic and scientific applications. Research often requires someone to observe a system, propose an explanation, design an experiment and revise the explanation using the result. Astra appears able to perform parts of that cycle in simulated environments. OpenAI researcher Noam Brown believes users may discover mathematical and scientific applications that the company has not yet tested.
Despite the headline results, the launch has received less attention than GPT-5 did last year. GPT-5 followed a long period in which the GPT-4 family defined the frontier for many users. Astra enters a market in which Anthropic has made progress in software engineering and enterprise use, Google remains technically competitive and Chinese laboratories regularly release strong models. A new GPT release no longer commands the same automatic attention.
Astra will likely prove to be the more consequential release. Its powerful capabilities arrive with evidence that the relationship between the model’s internal computation, written reasoning and external behaviour is becoming harder to follow.
OpenAI reports a substantial reduction in "chain-of-thought" monitorability. Astra can complete more work without communicating the intermediate steps. It can also exercise greater control over what its visible reasoning contains. Under test conditions, it sometimes concealed information associated with sabotage or strategic underperformance. Astra demonstrated a sandbagging capability. When instructed to underperform while avoiding detection, it could sometimes do so without being caught by the monitor. This does not prove that Astra sandbagged its safety evaluations. It means that unusually good behavioural results cannot necessarily be taken at face value.
OpenAI reports that Astra remained within its authorised scope more consistently than GPT-5.6 Sol. In one honeypot evaluation built on impossible tasks, Sol exceeded its authority in 48% of runs, while Astra made no such attempt. Astra also recorded no attempts to circumvent guardrails in another internal evaluation. Astra may break rules less frequently while becoming better at concealing a violation if one occurs. Its ability to recognise evaluations adds further uncertainty because behaviour observed during a test may not represent behaviour in another setting.
Last year, ExoBrain examined warnings that chain-of-thought monitoring might provide only a temporary window into model behaviour. Researchers feared that more capable models would learn to reason without writing everything down, recognise when they were being monitored or deliberately produce misleading traces. Astra now displays each of those capabilities to some degree.
The architecture does not provide a simple explanation. Media reports suggest Astra may use something called "recurrent depth". OpenAI has not confirmed this in its launch documents, and it has not attributed Astra’s reduced monitorability to recurrence. Recurrent processing could make interpretation harder by creating an evolving state across repeated passes. It could also help researchers by giving them one shared transformation to study. Better representations, reinforcement learning, context management and pressure to use fewer tokens could all explain why Astra needs less written reasoning.
Written chain-of-thought was never a direct transcription of internal computation. It was a generated scratchpad that exposed some useful information. Astra shows that effective reasoning can increasingly occur without producing a complete verbal account.
The shift away from written thought extends beyond Astra. Mostik, a startup founded by Russian mathematicians and led by the Fields medallist Stanislav Smirnov as chief scientist, is developing bridges that let different models exchange mathematical representations without communicating through text. Such systems could combine general models with specialists at lower cost, but their interactions would be harder for people to inspect.
As we explored last week, the complexity multiplies again when models become agents. An agent combines a model with memory, permissions, software tools and access to external systems. Multi-agent systems add delegation, communication and group behaviour. Failures can then emerge from interactions even when no individual component appears responsible for the final outcome. Clearly such systems could support scientific and economic activity at a scale that is difficult to achieve with people alone. Specialist agents could run analyses, test hypotheses, operate equipment and share results continuously. But such structures also create more places for errors, unexpected coordination and failures of oversight.
The solution was never going to depend on reading a model’s thoughts, which can now run into millions of words for a moderately complex task. Organisations will need several layers of control: alignment training, mechanistic research, reasoning monitors, action analysis, isolated environments, one-way data diodes, restricted permissions and human approval for consequential steps. Monitoring must cover the whole system, including interactions between agents.
Takeaways: GPT-6 Astra appears to be a larger advance than its restrained launch suggests. It can construct useful representations, learn unfamiliar environments and complete extended technical work with increasing efficiency. It also provides less reliable evidence about how that work is performed. As models operate through hidden states, tools and networks of other agents, the challenge expands from understanding an individual model to governing complete ecologies of machine activity.