Thinking Machines released Inkling, a 975-billion parameter mixture-of-experts model that activates 41 billion parameters during inference while supporting multimodal reasoning and a one-million-token context window.

The company made the model available through its Tinker platform for customization, alongside a smaller preview version for developers to test capabilities before scaling up.

Meanwhile, Perplexity launched SPACE, a sandbox platform designed to secure AI agents handling sensitive enterprise tasks. The system uses ephemeral sandboxes that self-destruct after task completion, with layered security including credential isolation and encrypted storage.

SPACE employs a Control Plane and Node-level Services architecture to manage credential access while enabling both on-premises and offline operations. The platform addresses growing enterprise concerns about AI agent security in production environments.

Anthropic moved closer to a potential IPO as investment banks Goldman Sachs, Morgan Stanley, and JPMorgan Chase began arranging investor meetings. The AI safety company recently closed a $65 billion funding round at a $965 billion valuation, surpassing OpenAI's $852 billion mark.

The potential public offering would test market appetite for AI companies amid the current boom, with Anthropic positioned as a potential bellwether ahead of OpenAI's own expected listing.

Research developments

OpenAI detailed GPT-Red, an adversarial model trained to generate attacks that expose vulnerabilities in frontier AI systems. The red-teaming approach reportedly reduced failures on prompt-injection benchmarks by sixfold for GPT-5.6 Sol.

Separate research demonstrated what researchers called the first experimental evidence of recursive self-improvement in AI systems. An autonomous research agent improved its own performance over eight days, designing novel search algorithms and reducing prompt sizes by 16x while building defenses against reward hacking.

NVIDIA expanded its Jetson Thor platform with Blackwell-powered T3000 and T2000 modules for edge robotics applications. The compact systems target visual AI and agent workloads, with early adoption from Amazon Robotics, Boston Dynamics, and FANUC already secured.