In mid‑2026 a series of high‑profile announcements showed that large AI models had begun resolving or disproving long‑standing problems in pure mathematics. Systems developed or run by OpenAI and Anthropic — and notably publicly accessible models such as GPT‑5.6 Sol Ultra and internal Astra family models — produced several surprising results over a few weeks. These events forced mathematicians and technologists to confront questions about attribution, verification, human roles in research, and institutional responses.
What happened and when
-
In January 2026 Daniel Litt (assistant professor of mathematics, University of Toronto) described contemporary models as roughly at contest‑problem level with room for improvement; by May–July the situation had escalated.
-
An internal OpenAI model reportedly produced a major advancement on the unit‑distance problem (originally posed by Paul Erdős in 1946), yielding “an infinite family of examples that provide a polynomial improvement” over the best known results and calling the consensus belief into question. OpenAI characterized this as the first time a prominent open problem central to a subfield was autonomously solved by AI.
-
On July 10, 2026 the public GPT‑5.6 Sol Ultra model returned a proof of the Cycle Double Cover Conjecture after being prompted to devote considerable effort; this set the tone for a hectic month of announcements.
-
In the following weeks several other old conjectures were reported as disproved or advanced: on July 20 a Jacobian conjecture disproof was announced in social posts associated with Anthropic (and reportedly a GPT model had produced the same counterexample earlier); July 23 brought a tweet about the Dinitz‑Garg‑Goemans conjecture being false; July 30 saw a claim about the Maxwell conjecture. On August 1, OpenAI published a blog post listing “Ten advances in mathematics and theoretical computer science” attributed to an internal Astra model family, and stated that reproducing the token usage would have cost roughly $2,000 at Sol API rates.
-
On August 4 a paper (or commentary) challenged OpenAI’s sourcing for an announced result (the non‑sofic group discovery), prompting OpenAI to revise some claims.
Reactions: praise, alarm, and a split narrative
Responses were sharply divided.
-
Some mathematicians called the achievements outstanding and saw them as a milestone that can reveal neglected corners of the field.
-
Others raised alarms about comprehension and trust. Terence Tao and others emphasized that AI outputs can be opaque, and that plausibly‑looking but incorrect proofs pose a serious problem.
-
Critics warned of ‘‘proof indigestion’’: if AI generates large volumes of candidate proofs and counterexamples, the human effort required to verify, formalize, and interpret them may become the bottleneck.
Why mathematics — and why are models good at it?
Several factors help explain why contemporary models succeeded so quickly in certain mathematical tasks:
-
Mathematics is highly verifiable. Logical correctness can, in principle, be checked formally, making it straightforward to shape and reward model behavior toward correct answers.
-
The mathematical literature is large, structured, and richly documented; the step‑by‑step reasoning of experts provides especially valuable training data.
-
AI systems excel at exhaustive search, cross‑referencing remote literature, and pursuing long, grind‑like reasoning chains that human researchers rarely attempt because of time and resource constraints.
These properties do not imply that mathematics is ‘‘easy’’ — only that its formal nature and the available data make it especially tractable for current architectures.
What AI still struggles to do: theory formation, abductive jumps, and understanding
Researchers from Google DeepMind and others have argued that while large language models can execute deductive proof from given premises, they are structurally weaker at the abductive ‘‘jump’’ required to propose new foundational premises or radically novel theories. Demis Hassabis has urged caution, noting that solving many Erdős problems does not equate to general artificial general intelligence.
Even where an AI produces a correct proof, the human role remains crucial: translating a machine output into conceptual insight, exposition, and application—the connective tissue that makes a result meaningful to the broader community.
Institutional and ethical challenges: attribution, data use, and verification costs
The new dynamic raises concrete policy and ethical questions:
-
Attribution: when an AI reuses or synthesizes existing partial work, how should credit be assigned between prior human authors, the model, and the institution running it? The Leiden University‑led Declaration on Artificial Intelligence and Mathematics (endorsed by several organizations and mathematicians) stresses that mathematics is not just a corpus of results but a community practice of understanding, and that attribution practices must reflect that nuance.
-
False positives: models can generate convincing but incorrect proofs; their low cost and high throughput make it harder to filter errors reliably.
-
Data and copyright: models often rely on published mathematical works collected under licensing conditions not designed for AI training, raising concerns about proper sourcing and legal compliance.
OpenAI responded to specific attribution concerns by saying that they helped prepare manuscripts, formalized proofs in Lean, and took responsibility for correctness while acknowledging the machine‑generated nature of the arguments.
Career and research implications
The economic and practical cost of research shifts when AI can cheaply explore many approaches (the OpenAI claim of roughly $2,000 of token cost was widely noted). Possible consequences include:
-
A move by some mathematicians toward higher‑level roles (synthesis, exposition, interdisciplinary connections) and away from pure problem‑solving.
-
A risk that younger researchers find mathematics less attractive as a career if machine systems capture the most visible breakthroughs and the social rewards that accompany them.
-
A reallocation of effort from discovery to verification and interpretation: more labor focused on checking, formalizing, and communicating AI‑generated results.
Net effect: neither apocalypse nor simple triumph
The mid‑2026 developments are not a binary judgment that mathematics is finished or that the Singularity has arrived. Instead, they expose a rapid transformation with mixed consequences. AI is very good at certain, well‑posed, verifiable tasks and at searching vast corpora for overlooked connections; but human judgment, conceptual insight, and the social processes of mathematical understanding remain essential.
The community now faces pragmatic choices: build protocols for verification and attribution, adapt training and career structures, and integrate AI tools in ways that preserve the epistemic goals of mathematics — clarity, understanding, and the capacity to pose new meaningful questions.
Per aspera, ad astra — through hardships to the stars. The models accelerate discovery, but whether those discoveries become usable, enlightening, and socially valuable depends on human institutions and work.
