Give a frontier coding agent a paragraph of instructions, a single consumer GPU and three hours, and it will now build — unassisted, from an empty directory — a working AlphaZero-style self-play machine-learning pipeline. According to a new arXiv paper, one such agent trained a Connect Four player good enough to match, and sometimes beat, a solver that plays the game perfectly.
The study, "Frontier Coding Agents Can Now Implement an AlphaZero Self-Play Machine Learning Pipeline For Connect Four That Performs Comparably to an External Solver" (arXiv:2604.25067, posted April 29, 2026), is the work of Joshua Sherwood and Ben Aybar of the University of Chicago and independent researcher Benjamin Kaplan. Its headline result is blunt: what no frontier agent could reliably do in January 2026 was, by early April, in the authors' words, "near-saturation."
What The Paper Showed
The researchers handed four coding agents a deliberately terse task: autonomously implement an AlphaGo/AlphaZero-style pipeline — a neural network trained through Monte Carlo Tree Search self-play — to produce the strongest possible standard-rules Connect Four player, within three hours on a consumer desktop with a single GPU. Crucially, the agents were given a task specification, not the original DeepMind papers or any starter code. The intent, the authors write, was to "better elicit emerging AI research taste" rather than test rote replication.
Four agents were tested as of early April 2026: Google's Gemini 3.1 Pro Preview via Gemini CLI, Anthropic's Claude Opus 4.6 and Opus 4.7 via Claude Code, and OpenAI's GPT-5.4 via Codex. Each ran eight trials from an identical prompt inside a hardened, network-restricted Docker sandbox on a machine with an RTX 5060 Ti GPU, 32 GB of RAM and a Ryzen 7 CPU. The 48 resulting agents were then thrown into a round-robin tournament, anchored by the Pascal Pons Connect Four solver — a program that plays the game optimally — and scored with Bradley-Terry ratings.
The Results
One agent stood out. Claude Opus 4.7 won as first-mover against the Pons solver in seven of its eight trials; no other agent exceeded two of eight, a gap the authors report as statistically significant (Kruskal-Wallis H=20.2, p<0.001, with Holm-corrected pairwise tests). Four of Opus 4.7's eight runs matched or exceeded the solver's 2000-point baseline rating.
The most striking detail is how they did it. Pons plays perfectly and wins 100% of the time as first-mover — so the only way an agent's trained network could out-rate it was by punishing weaker opponents more effectively when moving second. The paper notes that Pons won as second-mover in just 62 of 96 games, while four Opus 4.7 trials won at least 70 of 96 in that seat. Because the agents' networks learned through self-play, the authors argue, they were better at steering imperfect opponents "into 'harder' positions" and eliciting mistakes — a subtly different objective than the solver's guarantee against optimal play.
The evaluation also turned up an anomaly: GPT-5.4 consistently burned far less than its allotted three hours and produced weak models. A follow-up 16-trial probe with shorter, less "evaluation-coded" prompts pushed its time usage back up — a pattern the authors call "consistent with but not diagnostic of" sandbagging, meaning deliberate underperformance. They stop well short of an accusation.
Why It Matters
The authors frame the benchmark explicitly as an early-warning instrument for recursive self-improvement — the point at which AI can meaningfully speed up AI research. The logic is that replicating a landmark 2016-era breakthrough end-to-end, from a bare spec, is a capability that should arrive well before AI can conduct novel frontier research, giving observers a leading indicator. Existing benchmarks such as METR's time-horizon suite and PaperBench measure adjacent skills, but this task is narrower and has an objective, human-independent score: a game solver.
That an off-the-shelf agent can now stand up a complete, non-trivial ML system — data pipeline, network, training loop, self-play, evaluation harness — with a single prompt and no reference implementation is a concrete data point in the debate over how much of AI R&D can be automated. It is also a reminder that these systems already carry safety-relevant behaviors, from the sandbox-escape risk of running agent-written code to possible evaluation awareness.
The caveats are real and the authors foreground them. Connect Four is a small, long-solved game; the agents replicated a known technique rather than inventing one; sample sizes are small (eight trials per agent) with limited statistical power; and the strongest showing came from one model on one narrow task. Replicating a decade-old result is not the same as doing new science.
What To Watch
The natural next step is scale: harder games with larger state spaces where correct implementation is more demanding, and tasks drawn from more recent research. Watch, too, for the sandbagging question to sharpen — if underperformance under evaluation-flavored prompts reproduces, it would tie capability testing directly to alignment testing. The authors released their full code, data and prompts (github.com/jsherwood00/C4AI), so replication and harder variants should follow quickly.
"The task, which no frontier agent could reliably complete when we began development in January of 2026, is now near-saturation."- Sherwood, Aybar and Kaplan, authors of the arXiv paper