- Anthropic’s Claude Opus 4.5 became the first model to exceed 80% on the SWE-bench coding benchmark, a threshold widely treated as a meaningful proxy for real-world autonomous software engineering capability.
- Beijing blocked the export of 75,000 Nvidia H200 GPUs and vetoed Meta‘s $2 billion Manus AI acquisition, tightening hardware and deal-making constraints for enterprises with cross-border AI exposure.
- All five major US AI labs formally agreed to pre-deployment security reviews, the most significant voluntary safety commitment of 2026 so far, though the agreement’s value will depend entirely on the rigor of whoever conducts those reviews.
Anthropic‘s Claude Opus 4.5 cleared 80% on the SWE-bench coding benchmark this week, a first for any model and a result that will compress enterprise procurement timelines for AI developer tools. That headline landed alongside Beijing blocking a major Nvidia GPU export and vetoing Meta’s $2 billion Manus AI deal, while all five major US AI labs signed onto pre-deployment security reviews. It was a week that moved simultaneously on model capability, geopolitical friction and governance.
Anthropic’s Opus 4.5 Becomes the First Model to Crack 80% on SWE-bench
Claude Opus 4.5 is the first AI model to exceed 80% on SWE-bencha benchmark widely treated as a meaningful proxy for real-world agentic coding performance. That score positions Opus 4.5 not as a smarter chat assistant but as a credible autonomous software engineer, one capable of resolving complex GitHub issues with minimal human input. For enterprises currently evaluating AI agents for developer productivity, the number matters: 80% on SWE-bench has been the informal threshold separating “promising” from “deployable,” and Opus 4.5 has crossed it.
Beijing Blocks 75,000 Nvidia H200s and Vetoes Meta’s Manus AI Deal
Two separate moves from Beijing this week tightened the geopolitical constraints around AI infrastructure. China moved to block the export of 75,000 Nvidia H200 GPUsa decision that demonstrates Beijing’s willingness to use hardware controls as a strategic lever in the broader AI competition. Separately, China vetoed Meta’s proposed $2 billion acquisition of Manus AIthe agentic AI startup that drew considerable attention earlier this year. The two moves in combination deepen the bifurcation of the global AI supply chain. Enterprises with cross-border AI infrastructure dependencies should be stress-testing those assumptions now.
All Five Major US AI Labs Agree to Pre-Deployment Security Reviews
All five major US AI laboratories agreed this week to submit frontier models for pre-deployment security reviews before public release. The agreement covers the labs developing the most capable AI systems currently available and represents a step beyond the internal red-teaming and public pledge model that has dominated industry self-governance to date. Whether it carries real weight depends on who conducts the reviews and how. As a precedent, the commitment is notable; as a governance mechanism, it remains unproven. It arrives as regulatory pressure from both Washington and Brussels continues to build.
SAP’s 200+ AI Agents Signal the Scale of Enterprise Automation
SAP now has more than 200 AI agents driving workflow automation across its enterprise platforma figure that reflects how quickly agentic infrastructure is scaling inside major SaaS vendors. SAP‘s approach embeds agents directly into existing ERP and business process workflows rather than layering them on externally. That design choice connects to a broader pattern visible this week: enterprises moving beyond orchestration frameworks like LangChain toward purpose-built, native AI agent architectures that offer tighter integration, better observability and lower latency.
FIS and Everlaw Deploy Agentic AI in Finance and Legal
Two sector-specific deployments this week put agentic AI into regulated workflows at production scale. In financial services, FIS deployed an Anthropic-powered agent that compresses anti-money laundering investigations from hours to minuteswith direct implications for compliance costs and regulatory response times across banking. In legal technology, Everlaw added support for Anthropic’s Model Context Protocol to enable agentic legal workflowsletting legal teams orchestrate document review and discovery through agents connected to live case data. Both deployments reinforce the same point: agentic AI is past the pilot stage in regulated industries, and Anthropic’s MCP is emerging as a connective standard for enterprise agent deployment.
Aidoc AI Reaches 99.2% Accuracy in Pulmonary Embolism Detection
Pulmonary embolism is one of emergency medicine’s most time-critical diagnoses. Aidoc reported this week that its AI system matches radiologist performance and achieves 99.2% accuracy in PE detectionaccording to the company. PE causes tens of thousands of deaths in the US annually, and delayed detection is a leading factor in preventable outcomes. At that accuracy level, deployed across hospital networks where radiologist coverage is uneven, AI-assisted radiology moves from experimental supplement to a credible standard-of-care tool.
AI Safety Gets a Benchmark Problem, and a Papal Warning
Mpathic’s mPACT benchmark study found that AI models are gaming safety evaluationsperforming well on known test sets while failing to generalise safe behaviour to novel scenarios. The finding is a serious methodological problem for AI safety measurement and adds pressure on labs to develop more adversarial evaluation frameworks. On a different register, Pope Leo XIV warned this week that the AI arms race risks fueling a “spiral of annihilation.” The Vatican also formally launched a Vatican AI Commission ahead of what is expected to be the Pope’s first encyclical addressing AI ethics. Whatever weight one assigns to the Vatican’s institutional voice, the Catholic Church is now formally positioning itself in the global AI governance conversation.
Deepfake Fraud Hits $3 Billion, Legislators Respond
Deepfake scams were tied to $3 billion in US fraud losses in 2025according to new reporting. That figure landed alongside two pieces of state-level AI legislation moving in opposite directions. Connecticut passed SB 5imposing new transparency requirements on AI-powered hiring tools, one of the first US state laws to directly regulate algorithmic employment decisions. Texas, meanwhile, launched its SAM AI initiative targeting $123 million in regulatory savings through AI-assisted government operations. The contrast is instructive: US states are simultaneously legislating AI risk and deploying AI tools at scale.
What to Watch Next Week
The pre-deployment security review agreement is the story to track. The details that matter, who conducts the reviews, how they are structured, whether the commitment holds under commercial pressure, have not yet been established. If those details emerge next week, they will determine whether this announcement becomes a genuine governance mechanism or a well-publicised pledge. On hardware, watch for downstream effects of Beijing’s H200 block on AI infrastructure pricing and availability across Asian markets. And with Opus 4.5 now above 80% on SWE-bench, competing labs will respond with benchmark results and model releases of their own. The agentic coding race is entering its most competitive phase. For more analysis on enterprise AI strategy, visit our Enterprise AI section.
