On July 11, 2026, OpenAI announced that GPT-5.6 Sol Ultra had proved the Cycle Double Cover conjecture — a graph theory problem dating back about fifty years — in under an hour. The system split into 64 agents that attacked and critiqued each other's proofs.
A few days later, Tokyo-based physicist Yuji Tachikawa presented a string theory problem that had been stalled for six months to Claude Fable 5; the model produced a solution overnight and then wrote code to check its own answer. Both models and their outputs were made available to the public.
Practical implications
Previously, notable automated proofs could be run on closed, inaccessible models, which meant the scarcity of breakthrough results shifted from individuals to private labs. That protective explanation no longer applies: these recent examples were produced with models accessible to anyone with a login.
Equally significant is that the models went beyond proposing solutions and partly or fully verified them themselves — for example through a multi‑agent critique process or by generating code to validate results. This development potentially removes two scarcities at once: who can create discoveries and who can certify them.
Why it matters
If public models can rapidly produce and self‑verify solutions to hard theoretical problems, the distribution of discovery and the role of human certification may change. The concrete cases — GPT-5.6 Sol Ultra’s 64‑agent proof strategy for the Cycle Double Cover conjecture and Claude Fable 5’s overnight solution plus self-checking code for a stalled string theory problem — illustrate that these capabilities are now available under subscription or public access.
Next questions
Key follow‑ups include how the scientific community will assess and publish machine‑originated proofs and verifications, and what role traditional peer review and human expertise will retain. The new situation could substantially affect who benefits from and who validates future scientific advances.
