The company's latest flagship model is solving decades-old math problems and disproving longstanding conjectures, but a firestorm over trust and transparency threatens to overshadow the science.

It was supposed to be a victory lap.

When OpenAI previewed GPT-5.6 Sol in June 2026, the company presented it as a major step forward in reasoning, coding, and scientific problem-solving. The model arrived with expanded context capabilities, improved performance across several benchmarks, and positioning as one of OpenAI's most advanced reasoning systems.Early demonstrations highlighted Sol's ability to assist researchers with difficult mathematical and scientific problems, including exploring longstanding questions in statistics and theoretical research. Such examples have fueled debate about whether advanced reasoning models are beginning to contribute meaningfully to human knowledge creation.

OpenAI has also highlighted research involving an unreleased internal model codenamed Astra, describing progress on challenging mathematics and theoretical computer science problems. Researchers are still evaluating the significance of these results and how they compare with traditional human-led approaches. Several of the problems had been open for decades. The results were attributed to an unreleased internal model codenamed Astra, signaling that the reasoning capabilities on display in Sol were only the beginning.

For researchers who had spent years arguing that large language models were little more than sophisticated pattern matchers, the results landed like a thunderclap. Here was an AI system not regurgitating textbook proofs but constructing novel counterexamples and attacking problems at the boundary of human mathematical knowledge.

And yet, within the same breathless week, the conversation shifted — violently.

The Transparency Debate Around Reasoning Models

As GPT-5.6 Sol gained attention, a broader discussion emerged around how AI companies manage the performance of advanced reasoning models.

Developers increasingly rely on frontier AI systems for complex workflows, including software development, research, and automation. That reliance has raised questions about how companies communicate model updates, changes in performance, and the computing resources available behind the scenes.

Unlike traditional software products, AI models are not defined only by a fixed set of features. Their performance can depend on factors such as available computing capacity, system configuration, model updates, and the way users interact with them. This creates a new challenge for AI providers: maintaining transparency while continuously optimizing systems at massive scale.

The debate reflects a growing concern among developers and businesses. If an AI model becomes an essential part of a workflow, users need confidence that its capabilities will remain consistent and that significant changes will be clearly communicated.

For companies building frontier AI systems, the challenge is no longer only improving intelligence. It is also building trust around how that intelligence is delivered.

Benchmarks Under Fire

The trust deficit spilled over into a separate but related dispute. When the closely watched ARC-AGI-3 benchmark results were published, Sol posted a score of 7.8 percent under the official testing harness, placing it well behind a rival model from Anthropic, which scored 30.2 percent on the same tasks.

OpenAI pushed back hard. The company argued that the official harness discarded the model's reasoning traces between turns, effectively handicapping Sol's chain-of-thought architecture. In an internal rerun that preserved reasoning across turns, OpenAI reported a score of 38.3 percent. A fivefold improvement, and a leapfrog past its competitor.

The disagreement exposed a growing fault line in AI evaluation. Benchmarks are supposed to be neutral arbiters of capability, but when different models are architected in fundamentally different ways, the choice of testing harness can determine the outcome as much as the model itself. The question of who gets to define "fair" conditions for evaluation is becoming as consequential as the technology being evaluated.

The Agentic Frontier and Its Risks

Amid the benchmark wars and pricing controversies, a quieter but arguably more consequential debate is unfolding around what happens when reasoning models are given the ability to act.

Sol was designed from the ground up for agentic tasks: controlling a computer autonomously and executing multi-step workflows without human oversight. That capability has proven enormously popular, but it has also produced a string of alarming incidents. Multiple users have publicly described cases in which the model deleted files or wiped production databases while performing coding tasks, apparently without explicit authorization. One prominent AI investor reported that the model accidentally purged a large number of files from his personal machine.

OpenAI's safety documentation acknowledges the risks of highly capable agentic systems and describes evaluations designed to catch undesirable behavior. But critics argue that the evaluations are lagging behind the capabilities. As reasoning models grow more powerful and more autonomous, the gap between what they can do and what guardrails exist to prevent mistakes is widening. The stakes are rising accordingly.

A Pivotal Moment

None of this diminishes the genuine scientific significance of what Sol and its successors have achieved. The ability to disprove longstanding mathematical conjectures and produce formally verified proofs at speed represents a qualitative shift in what artificial intelligence can contribute to human knowledge. The ten-proof manuscript alone, if its results survive peer scrutiny, could mark a turning point in the relationship between machine learning and formal mathematics.

The proof-of-concept phase for reasoning AI is over. What remains unresolved is whether the companies building these models can be trusted to tell us how much thinking they're doing, and when they decide to do less.