OpenAI's GPT-6 Astra posts benchmark and adoption wins as Meta's Muse and Cognition's SWE-2 expand the agent landscape

OpenAI's GPT-6 Astra led ulam.ai's ErdosBench math benchmark (106 of 226 problems) and showed a 'step change' in spatial reasoning on StationeryBench, while new OpenAI case studies detail production use at Cognition and Perplexity. Meta launched its…

OpenAI's GPT-6 Astra dominated this week's news. On ulam.ai's ErdosBench it scored 3.23, solving 106 of 226 open math problems (43 fully, 27 disproven), ahead of GPT-5.6 Sol's 3.12 and 78 solved at maximum reasoning; chief scientist Jakub Pachocki said OpenAI deliberately skipped targeted math optimization to prioritize recursive self-improvement and automated alignment research. Early StationeryBench robotics results over 200 trials showed Astra fully completing 7 of 100 dual-arm YAM robot tasks (median progress 46/100) versus zero for Ai2's MolmoAct2 (12/100), which researcher Yoav Artzi called a 'step change in spatial reasoning.' OpenAI also published its first Astra customer case studies — Cognition uses it across Devin's cloud agent, CLI, and desktop for automated testing, and Perplexity uses it via API for communications, production edits, and end-to-end tests — plus developer guidance to shorten skills, trim AGENTS.md requirements, and define 'done,' since Astra may stop earlier than GPT-5.6 Sol. Meta launched Muse, a WhatsApp-controlled agent running on an isolated cloud VM with a Sentinel credential gatekeeper and Stripe Link one-time-card payments; its model reportedly scored 44-48 on Artificial Analysis Intelligence Index v4.3, near GPT-5.6 Sol's 47, but The Verge's hands-on found it surfacing Instagram/Facebook API-derived interests beyond users' visible ad-topic settings, which Meta disputes is a sharing issue. Cognition's SWE-2, post-trained from the 2.8T-parameter Kimi K3 MoE, posted vendor-reported scores of 50.0% on FrontierCode 1.1 Main and 92.8 on Terminal-Bench 2.1, within one point of Claude Fable 5.1 at a claimed 64% lower cost, though it trails Fable 5.1 and GPT-6 Astra on long-horizon Terminal-Bench 4.0 tasks. Separately, a Hugging Face paper found six frontier models from four labs converging on one imagined architecture under school-audience framing, coining 'epistemic jailbreak.'

• GPT-6 Astra scored 3.23 on ulam.ai's ErdosBench, solving 106 of 226 open math problems (43 fully, 27 disproven); GPT-5.6 Sol trails with 3.12 and 78 solved at maximum reasoning.
• Benchmark developer Przemek Chojecki estimated Astra's gain at 5-10% across tested math-research skills; OpenAI chief scientist Jakub Pachocki said targeted math optimization was skipped in favor of recursive self-improvement and automated…
• Mathematician Terence Tao warned at the 2026 International Congress of Mathematicians that AI-generated proofs could shift mathematics from proof scarcity to proof overload.
• On StationeryBench (five desk-object tasks, identical dual-arm YAM robots, 200 trials), Astra fully completed 7 of 100 tasks with median progress 46/100; Ai2's MolmoAct2 completed zero with 12/100.
• Cornell/Google DeepMind researcher Yoav Artzi called Astra's result a 'step change in spatial reasoning,' said it approaches human-level accuracy on the unpublished REMAP benchmark, and speculated training on large 3D datasets such as…
• Cognition uses GPT-6 Astra across Devin's cloud agent, CLI, and desktop: Devin tested the iPhone game Otter Run (returning a simulator recording plus coverage report) and fixes customer-reported bugs from screenshots with verification;…
• Perplexity cofounder and Chief Strategy Officer Johnny Ho says GPT-6 Astra crafts communications, edits production systems, and builds mock services to test applications end to end, with far less supervision than previous models.
• OpenAI's Eric Provencher recommends shorter skill descriptions (too many or conflicting ones truncate in Codex's context), selective AGENTS.md document reads, explicit permissions for safe operations like local test runs, and upfront…