Harvard's amazing Google study: GPT-6 Astra has solved 98% of the world's physics problems!

  It's the first time in the history of the world.

It's so shocking.

In fact, a recent study has shown that the number of people living with HIV/AIDS in the United States has decreased.GPT-6 Astra has solved 98% of all new theoretical physics problems！

This top-level theoretical physics evaluation of TheoPhysEval has just shocked the academic community.

These unpublished front-end puzzles take top experts an entire day to solve.

But the GPT-6 Astra not only scored near the top of the table, it even produced a calculation error by experts on the subject!

Human experts admitted that the GPT-6 Astra's score dropped from 88% directly to 98%.

Unsurprisingly, MIT's latest findings show that in the past two years, the ability of cutting edge AI to solve theoretical scientific problems has skyrocketed 17 times!

After mathematics, quantum field theory, cosmology, and economics are being taken over entirely by AI.

Today's big models are extremely difficult to evaluate, as the online puzzle libraries have long been "eaten" by them.

To explore the real reasoning behind AI, physicists from the world's top institutions have worked with AI to dig a crater called TheoPhysEval.

The institutional line-up involved in the topic is very luxurious, including:He has been a visiting scholar at the International Centre for Theoretical Sciences (ICTS), the Tata Foundation Institute (TIFR), the University of Texas at Austin, the University of Tokyo, Harvard University, the University of Pennsylvania, Duke University, and Google DeepMind.

The project contributors include dozens of renowned scholars such as Suvrat Raju, Sudip Chakraborty, A. Ramesh Chandra.

In order to prevent AI from cheating,The topics are all from the latest research notes not published by the experts.

These 50 hard-core puzzles cover quantum field theory, holographic theory, and so on. Even top experts in the field of division can take a day or even days to solve a problem.

Even the world's top experts in this segment of the field would take a day or more to solve one without AI assistance; it would take a very long time to solve it, even if replaced by experts in a neighboring field, such as physics giants.

The tests began. The researchers gave the 50 questions to GPT-6 Astra, Opus 5.5, Gemini 3.8 Flash, and the open source DeepSeek-V4.1-Flash, among others.

As a result, GPT-6 Astra initially scored around 88%, and Claude Opus 5.5 scored 92%.

But the next thing to come came as a shock to everyone. Project researcher Nirmalya Kajuri revealed: "Astra is extremely stubborn when it comes to examining errors".

It not only gives complete guidance, but also insists on the chain of thought:The logic itself is unmistakable, and the standard answer given by human experts is wrong!

In real-world test tracks, the Astra ended up crushing the entire course with an astonishing 98% victory rate.

In the face of AI's "facial output", experts can only rewrite old drafts.

The truth is staggering: in a number of topics, experts have found themselves buried deep in the steps of complex computation, and there are indeed extremely subtle omissions - such as symbol errors, approximate condition errors, etc.

And AI's guiding principle is that it's really rigid, perfectly seamless!

The standard answer was subsequently revised by experts "Red Faces".

After the revision, GPT-6 Astra's score rose to a high of 98% in 50 questions. They solved 49!

This time, AI not only did the right thing, but also gave human scientists' answers to "reverse criticism".

You might be wondering, what are the mysterious physics puzzles that make it difficult for top scientists to solve all day long?

Appendix A to the paper reopens as if it were opening a bookstore on the front lines of human physics.

Below are some of the challenges that Astra has successfully mastered in this assessment.

Black-string Page time calculation: Black-string Page time

This is a pioneering study of the information paradox of black holes.

The topic requires AI to consider a black-string background containing two membranes in the anti-desicter space (AdS4) and AI needs to compare the area of different extreme curves in the AdS4 space to calculate Page time.

Astra not only identified the five-dimensional time-space equations in the literature, but also independently extended them to three-dimensional adaptations, and hand-written Python scripts processed the odd-point differential equations to obtain perfect numerical values.

All-electron atomic solver: all-electron atomic solver

The title requires AI "from the ground up" to write a self-contained full-electron atomic structure solver to calculate the orbital energy of hydrogen and hydrogen at base states.

Astra uses the local density approximation (LDA), the Dirac-Slate exchange function, and even calls for VWN5 parameterization.

Starting from scratch, it worked out the code for solving the Kohn-Sham equation, and calculated lithium and silicon-based orbital energies that were in perfect alignment with NIST's micro-Hatteries benchmark data in the US!

Axis conversion rate: Axion conversion rate

This is a popular direction for searching for dark matter in high-energy physics. It requires the rate at which the axis is converted into photons by scattering electrons in a specific intermediate constant medium in the absence of an external electric field.

Faced with complex intermediate electrical constants and conjugation conditions, Astra perfectly substitutes the natural unit system (ħ = c = 1), zero-error formulates the rate at which the axis is converted into photons, and calculates the exact floating-point number.

Chern-Simons β function: Chern-Simons beta function

If you knew how easy it is to get your hair out of your hair, you'd be amazed at what Astra did.

This topic examines the contributions of the three-dimensional Chen-Simon set of norms to the theory of matter, involving complex nonlinear objections to quantum representation and infinite reintegration.

Faced with the difficulty of this three-dimensional quantum field theory, Astra accurately identified the transmitter and the vertex, and completed a complex very small deduction method to give a closed-form coefficient of the β function.

In addition, Astra has successfully addressed the following issues:

Disturbance of boundary layers in fluid mechanics in the direction of scale computation;

solving the central charge of a one-dimensional D-type fermionic system in condensed matter physics;

Distribution reintegration calculation of the two-circle quadrants of the scattering amplitude in D = 6 dimensions in Central European Geometers;

Seiberg's theory of parity and 14 node oscillations, which deals with the supersymmetric symmetry of the field in quantum field theory;

In string theory, the IIB type supergravity is calculated as the magnitude of a wormhole in Euclidean geometry under six-dimensional ring confinement.

It is hard to imagine that Astra could have conquered these divine puzzles and swept through the barrier of human intelligence.

TheoPhysEval uses a rigorous testing mechanism.

In order to simulate real scientific research, the test is divided into two tracks.

Real trajectory:It allows AI-connected searching, self-writing code calling tools.

This is the first time I've seen this.It's a pure disconnect that only tests AI's inner workings.

As a result, even in a disconnected orbit, both Astra and Opus 5.5 have scored 92% higher!

Obviously, the front-end model isn't just a search engine.They are in the pre-training phase, have internalized the literature of anthropophysics and have a deep understanding of key formulas and underlying logic!

What's more, in order to prevent AI from blurring in formats, subject matter experts have used "digital hacking techniques" with extreme reluctance.

To prevent model blindness, it is mandatory to convert all complex parsing expressions into extremely high-precision numerical results with interference items.

The subject will be forced to substitute some cold, meaningless floating-point numerical interference, requiring AI to eventually output an accurate decimal number string.

This fundamentally eliminates the possibility of AI blinding, formulating, or formatting complexity. The only figures that can be obtained show that AI's guidelines and programming are 100% correct.

The biggest threat to global research institutions is the cost of computing.

To solve such a world-class physical problem, the API Token cost of GPT-6 Astra is:On average, less than $3 per question!

For a cup of coffee, you can hire a never-ending academic worker. The traditional "human-sea tactics" have failed.

TheoPhysEval is just a shorthand for the field of physics. A similarly published study by MIT's Neil Thompson team directly alerted the theoretical science community.

They tried to answer the question:How fast is AI evolving in the entire field of theoretical science?

To ensure absolute fairness, the MIT team extracted the theoretical difficulties from their most recent papers published between December 2025 and May 2026, and tested the entire system model of o1, o3 and GPT-5 in a disconnected network environment.

The field of testing almost covers the crown of human intelligence:Probability theory, statistical theory, economic theory, high energy physics, gravity and cosmology, and quantum physicsWait a minute.

These are heavily reliant on theoretical derivations, without the need for experimental verification of the results.

The test data is incredibly shocking.

In mid-2024, the best AI in these fields will have a problem solving rate of only 2%.

Just two years later, in mid-2026, the resolution rate for front-end AI jumped directly to 35 percent, turning the whole world upside down.17.5 times！

Moreover, this index-level growth curve shows no signs of slowing down. After OpenAI announced last month that it had solved the Millennium Grand Prix puzzle in mathematics, panic officially spread to physics and economics.

Thompson's half-hour internal speech at MIT was silent.

Is AI so violent that human scientists are out of work?

There is a saying in physics: "Shut up and calculate!"

The paper says directly:「It is no longer appropriate to assign this type of validating calculation to young researchers.」Because AI is doing better, more accurate, cheaper.

Human "taste" has become immeasurably important.

In the entire test, Astra also submitted a white paper on the 1b question: reconstruct autonomous differential equations from orbit.

The question gives a curve with noise, requiring an inverse differential equation.

As a result, all top-level AIs have been overwhelmed by this need for flexible change.

This is because AI does not "turn around" and has no physical intuition. When information is asymmetrical or requires reverse thinking, AI often gets stuck in a dead end.

In the end, it takes humans to hold the steering wheel.

The future of scientific research will be "multi-cycle human-machine interaction". Humans provide inspiration, direction and academic "taste", and AI is responsible for guiding and mass computing.

The nature of dark matter, which has plagued humanity for hundreds of years, may be under attack in the coming years.