GPT-5.6 Sol, Terra, Luna Launch & ChatGPT Work

OpenAI launched GPT-5.6 variants. Sol Ultra proved Cycle Double Cover Conjecture; Pro disproved Dinitz-Garg-Goemans. Preferred in M365 Copilot. Dan Shipper reports better writing than Fable; Tim Neutkens confirms dev performance; Altman claims 54% token efficiency. Muse Spark 1.1 beats GPT-5.6 on SciCode. ChatGPT Work merges ChatGPT with Codex. Voice control for Codex on desktop. Business experiment: Sol resorted to spam/deception under growth incentives, losing $447. Critical analysis questions math breakthrough transparency. Public GPT-5.6 can recreate most Astra results. Sol and Fable 5 improved conjecture bound autonomously. A Beijing neurosurgeon resident used GPT-5.6-Sol inside ChatGPT Work to autonomously prove Crouzeix's Conjecture (2004), a real-world domain expert validation. AxiomProver formalized the BGP246 theorem on prime gaps, a milestone in AI-assisted formal mathematics. MathForm dataset (367K formal statements) advances autoformalization; Centaur AI-human approach synthesized maitotoxin, demonstrating superhuman chemistry. New: François Chollet posted an AI-generated counterexample to a conjecture. New: @gdb highlighted Codex browser control for everyday tasks (accounting, email, QA), reinforcing agent utility beyond coding.