Central to Laguna S 2.1’s design is a context window of up to one million tokens in both thinking and no-thinking modes, allowing entire problem-solving sessions to remain in memory. In thinking mode, the model dedicates large token budgets to internal deliberation, exploration, and backtracking while still preserving prior steps, enabling complex tool-assisted workflows to unfold coherently over extended periods. This configuration directly targets tasks where conventional short-context systems drop intermediate results, keeping the full reasoning trajectory available for inspection. In benchmark reports, this 1M-token context is paired with strong long-horizon coding performance, including a 70.2% score on Terminal-Bench 2.1 in thinking mode.
Erdős Problem #397, posed by Erdős, Graham, Ruzsa, and Straus in 1975, is a combinatorics and number theory question centered on constructing infinite structures with delicate arithmetic properties. It has long been treated as a challenging open-style construction task rather than a routine exercise, serving as a benchmark for work in additive combinatorics. A modern AI-assisted solution was first credited to GPT-5.2 Pro in January 2026, establishing a precedent for using frontier reasoning models to attack decades-old mathematical problems.
Poolside states that Laguna S 2.1 independently rediscovered a proof of Erdős Problem #397 despite its November 2025 knowledge cutoff, which predates the GPT-5.2 Pro result and prevents direct imitation. Over roughly 68 minutes of continuous operation, the model acted as an agent, issuing tool calls and writing Perl code to carry out number-theoretic calculations that shaped its search among candidate constructions. Instead of a single opaque answer, it produced a complete reasoning trajectory that iteratively refined hypotheses, discarded dead ends, and culminated in an explicit construction generating an infinite family of solutions.
Poolside presents the run as a benchmark for long-horizon constructive proof search, emphasizing that the published trajectory allows researchers to inspect how the model navigated the combinatorial landscape. The ability to sustain such a search inside a single one-million-token context while coordinating code execution with symbolic reasoning illustrates why Laguna S 2.1 is framed as an agentic coding model rather than a static code generator. This approach reflects the broader trends in AI world models that enhance predictive capabilities for complex tasks.
Reported results on DeepSWE v1.1, including its score in thinking mode, support the view that controlled internal reasoning is central to its performance on difficult multi-step tasks.
Economically, the experiment highlights the impact of sparse Mixture-of-Experts design and low-cost deployment. Poolside reports that the full 68-minute run, including tool use and internal deliberation, can be executed for less than ten cents of API spend. By pairing high-end mathematical reasoning with commodity-level pricing and single-machine deployability, Laguna S 2.1’s solution to a decades-old combinatorial problem illustrates how open-weight MoE systems may begin to democratize access to advanced automated problem solving.
