GPT-6 Astra Solved a Competitive Programming Problem by Leaving Competitive Programming Behind

In 2019, a comment appeared under the Luogu problem P4484. It read, “orz EI!” That was a tribute to Elegia, the problem setter. Seven years later, a model called GPT-6 Astra produced a new solution. It did not enumerate integer partitions. It pulled Plancherel measure from analytic combinatorics, combined the Baik-Deift-Johansson theorem with the Painlevé II differential equation, and reduced the complexity to polynomial time.

Contestants on the problem page reacted the same way. This material is not taught in OI. It is not taught in math contests either.

Vals AI’s IOI benchmark provides numbers. From 2024 to 2026, Astra scored 100 percent. GPT-5.6 Sol, the next model, scored 91.17 percent. The test conditions matter. The agent saw the statement, the sample grader, and one sample. It had no test data, no internet, and no submission feedback. Astra also scored 100 percent on 2026 problems that appeared after its training cutoff. On FrontierMath Tier 4, it scored 97.6 percent. With explicit chain-of-thought disabled, it handled tasks equivalent to about 30.9 minutes of human work. The previous model, GPT-5.6 Sol, covered about 3.6 minutes.

Those numbers describe the ceiling. The unsettling part is how Astra solves the problems.

In P4484, Astra switches from discrete enumeration to analytic representation. The Robinson-Schensted correspondence maps a permutation to a Young table. The LIS length equals the first row length. The problem becomes a sum over integer partitions. Contest solutions stop there and enumerate partitions. Astra continues. Under the Plancherel measure, the edge distribution of a random Young table is equivalent to the Tracy-Widom distribution. That distribution connects to the Painlevé II differential equation. Along this path, the hook-length product becomes a determinant. Sparse matrix high-order difference acceleration and the finiteness of D-finite polynomials turn exponential enumeration into polynomial analytic computation.

The important part is not that Plancherel measure or Painlevé equations are obscure. They are mature topics in the literature. The important part is the combination. Astra strings together RSK correspondence, Plancherel measure, Frobenius-Schur scalar products, and D-finite function theory into one computational path. Doing that requires understanding representation theory, random matrix theory, and analytic combinatorics at once. It also requires knowing which transformation to apply at which node. That cross-domain wiring is the striking part.

P8352 is another case. The standard solution is DP套DP, a two-dimensional knapsack transition with complexity around O(n^3 k^4). Astra uses Karatsuba algebra to avoid the two-dimensional convolution. Karatsuba is known for big integer multiplication, with complexity O(n^1.585). Using it to speed up convolution inside a DP transition requires deep knowledge of memory access patterns and computation graph structure. This is operator-level engineering optimization. One commenter put it well: this kind of optimization is far beyond OI, and many algorithm engineers do not know these operator methods.

Astra’s ability is combination optimization inside the convex hull of human knowledge. It does not invent new mathematics. Plancherel measure, Painlevé equations, D-finite functions, and Karatsuba all exist in the human knowledge base. Astra retrieves the most relevant combination and checks that the combination is mathematically consistent. That process is hard. A human expert cannot master representation theory, random matrix theory, analytic combinatorics, and operator optimization at the same time. When a problem crosses these fields, a human usually needs years of interdisciplinary training to build the necessary intuition. Astra does the retrieval and combination in minutes.

This is still an open-book exam. The prerequisites are a clearly defined problem, a verifiable correctness standard, and an explored knowledge space. P4484 has a clear statement: find the expected LIS length modulo 998244353. Correctness is fixed. Astra’s job is to find the fastest computational path inside that frame.

LiveOIBench adds another data point. The study found that stronger models do better not because they explore many random ideas, but because they invest more effort in focused problem analysis and verification. Models still perform worst on problems that need a clever new scheme, especially when the solution depends on breaking the task into many interrelated subproblems. That distinction matters. Astra is good at deep combination within existing approaches, not at creating new approaches.

The impact on algorithm engineers is less simple than a headline about job loss. Astra’s performance on P4484 depends on a fully defined problem. Input, output, correctness standard, and performance metric are all fixed. Under those conditions, Astra can use its entire knowledge base to find an optimal solution.

In real algorithm engineering, defining the problem is often the hardest part. A 2026 industry analysis put it this way: the core competitiveness in the AI era is shifting from doing things right to asking the right questions, defining the right metrics, and verifying the right causal path. That is the weak spot of pure statistical models. Another source said most AI projects fail not because of the model, but because the problem framing is wrong.

The hiring market reflects this asymmetry. The supply-demand ratio for high-performance computing engineers is 0.15, meaning one job seeker for seven companies. For SLAM algorithms and planning-control algorithms, the ratios are 0.21 and 0.23. An algorithm engineer at a Hangzhou AI company described the change: three years ago, his core work was hand-coding. Now 90 percent of his work is directing multiple AI agents.

The keyword is directing. The role is shifting from implementer to orchestrator. The engineer defines optimization goals, designs the search space, and verifies AI output. Implementation ability still matters. To judge whether AI output is reliable, an engineer needs to understand the details. When Astra says it reduced complexity to O(n^{23/12} log^2 n), an engineer without the relevant math background cannot verify the claim. When it says Karatsuba algebra accelerates a two-dimensional convolution, an engineer without memory access knowledge cannot judge whether the optimization works on specific hardware.

Back to P4484. Elegia’s intent, filtering the ability to find an optimal solution under constraints, still works. The carrier changed. AI did not make algorithm engineers extinct. It redrew the boundary between algorithm and engineering. On one side are defining the problem, designing constraints, and verifying correctness, which still need human judgment. On the other side is searching within a given problem space, which AI is taking over. Astra’s P4484 solution is a signpost. It marks the current boundary of AI ability: combination optimization inside the known knowledge convex hull. Outside that boundary remains human territory.

The problem still uses modulus 998244353.