AI Enters the Execution Era as Models, Agents, and Verification Converge

Frontier Models Move Beyond Generation

Announcements on September 8 and 9 showed AI systems taking on increasingly complete workflows. TechCrunch reported that an unreleased OpenAI model generated a proof related to the Navier–Stokes problem through roughly 10,000 coordinated agents over 88 hours, followed by formal Lean verification. OpenAI also released ChatGPT Images 2.5 with faster generation, sketch completion, templates, and regional editing. Inception reported that Mercury 2.5 reaches 1,107 tokens per second with a 260,000-token context window.

Open Models Reshape Cost and Deployment

Community benchmarks showed a 2.8-trillion-parameter Kimi K3 mixture-of-experts model running from four SSDs on a 128 GB M5 Max MacBook Pro at about one token per second. Perplexity launched Portable Computer, a local agent built on Alibaba’s open Qwen3.8-27B and optimized for NVIDIA DGX Spark. DeepSeek also announced lower Flash-model prices, including a 60% cache-hit input reduction during off-peak hours.

Capital Follows the Agent Economy

Mistral closed a €3 billion Series D at a reported €21 billion valuation, while Cognition reached a reported $48 billion valuation. Qualcomm and Amazon announced a multi-generation partnership covering data-center chips, inference, and optical interconnects. Meta’s Muse launch illustrated the personal-agent layer, but internal testing also highlighted the need for authorization, monitoring, and reliability controls.

Security and Evidence Become Product Requirements

Anthropic confirmed that information-stealing malware had been used to hijack Claude sessions and consume account quotas. Google’s Threat Intelligence Group reported that TeamPCP used an autonomous multi-agent framework to harvest credentials after supply-chain compromises. OECD PISA results further suggest that AI works best as scaffolding: students using it once or twice weekly outperformed both lower- and higher-frequency users, while evaluation-focused classrooms performed better.