- Claude Opus 4.5, released in November 2025, hit 80.9% on SWE-bench with 30-minute autonomous operation, while Claude Sonnet 5 cut agentic development costs by 86%, accelerating enterprise adoption decisions.
- The EU AI Act’s transparency rules are now enforceable, exposing non-compliant AI deployments to fines up to €15 million, with ChatGPT, Reddit and Roblox simultaneously designated under the DSA’s strictest oversight tier.
- A €387.8 million AMD-only European supercomputer procurement and enterprise survey data showing a 14-point preference for non-Nvidia chips put direct pressure on Nvidia’s full-stack lock-in strategy.
Anthropic posted its strongest technical week in months, the EU moved AI regulation from warning to penalty phase, and Europe’s largest AI procurement decision pointedly excluded Nvidia. Three developments, each consequential on its own; arriving together, they compress what would normally be a quarter’s worth of enterprise AI decision-making into a single news cycle.
Claude Opus 4.5 and an 86% Cost Drop
Claude Opus 4.5, released in November 2025, was the first model to clear 80% on SWE-bench Verified (80.9%) and could sustain 30 consecutive minutes of autonomous operation, a milestone at the time for enterprise workflows like code review passes and multi-step data pipeline runs. Anthropic has since extended that trajectory through several more releases, including Opus 5. Separately, Claude Sonnet 5 cut agentic AI development costs by 86%, removing the economic objection that has kept many teams watching agentic architectures rather than deploying them. The two releases together push Anthropic into direct contention with OpenAI for enterprise agentic workloads at exactly the moment procurement teams are choosing platforms to standardise on.
EU Enforcement Goes Live
European regulators moved from policy to enforcement. The EU AI Act’s transparency provisions are now active, with opaque AI systems facing fines up to €15 million for failing disclosure requirements. Companies deploying AI in high-risk categories must now demonstrate clear explainability to regulators or face penalties, not warnings. At the same time, the European Commission designated ChatGPT, Reddit and Roblox under the Digital Services Act’s strictest oversight tier, adding mandatory algorithmic audits, crisis response obligations and researcher data access requirements to their compliance burden. Any AI company with material European user exposure is now operating inside one of the world’s most demanding regulatory environments, and the cost of non-compliance is concrete.
Nvidia’s Lock-In Strategy Faces Pushback
The week’s chip story had two sides. Nvidia revealed that its $99 billion investment strategy is explicitly designed to achieve full-stack control across chips, networking, software frameworks and developer tooling. The logic is straightforward: make switching costs prohibitive. The market’s response, also this week, was instructive. Europe committed €387.8 million to an AMD-only supercomputer, a procurement decision that signals European institutions are treating supply chain diversification as a policy priority. A separate enterprise survey found that enterprises now favour non-Nvidia AI chips by a 14-point margin. The lock-in strategy may be accelerating the diversification it was designed to prevent.
Google’s Velocity Becomes Its Own Problem
Google had a broad week. Its TPU and software stack addressed up to 75% of AI memory costs, according to the company, strengthening the case for Google Cloud in cost-sensitive deployments. The model side told a different story. Three Gemini Flash releases within six weeks are forcing enterprise teams into urgent migration decisions before current integrations are deprecated. Release velocity is a competitive signal until it becomes an operational burden for the customer base. Separately, a $400 million DeepMind acquisition now drives Google’s agentic Gemini push, bringing world-model research directly into contact with production infrastructure at a scale few competitors can match.
Security: Prompt Injection and a $20 Safety Removal Service
Researchers demonstrated the “GitLost” prompt injection attack, capable of exposing private GitHub repositories by manipulating AI coding assistants into leaking repository contents through crafted inputs. As AI agents gain access to sensitive codebases and file systems, prompt injection becomes an enterprise-grade attack surface. The week’s other security story sat at a different extreme: Abliteration.ai is openly selling permanent AI safety guardrail removal for $20 a month, stripping alignment training from models and offering uncensored outputs as a subscription. The service poses a direct challenge to EU AI Act enforcement: commercial safety removal was not a hypothetical when those rules were drafted.
Infrastructure Costs, from Grid to Desktop
AI hyperscalers are now building their own power generation capacity as grid connection delays threaten to bottleneck data centre expansion. In India, HyperVault announced a 1GW AI data centre campus, confirming the infrastructure race extends well beyond the US and Europe. Closer to the enterprise budget line, an analysis found that Microsoft, Slack and Zoom AI add-ons can run hundreds of dollars per seat annually, a figure now prompting CFOs to demand ROI evidence before approving blanket seat rollouts.
Healthcare, Public Spending and AI-Generated Pizza
UCSF integrated OpenAI’s ChatGPT directly into its Epic electronic health record system to cut physician documentation time, one of the most significant clinical AI deployments at a major academic medical centre to date. Wellington ratepayers, meanwhile, funded a $435,000 Deloitte report that the city’s mayor publicly claimed was written by AI, triggering immediate public scrutiny over cost and transparency. And Uber Eats caught five separate pizza restaurants using an identical AI-generated food photograph on their listings: a small but concrete illustration of how AI-generated content erodes consumer trust when the duplication surfaces.
What to Watch Next Week
With the first €15 million fine threshold now active, watch for the first named EU enforcement actions; the shift from warnings to penalties will set a compliance tone that extends beyond Europe. On the model front, whether OpenAI responds to Anthropic’s Opus 4.5 benchmark claims with a counter-release or a public rebuttal will shape the agentic capability narrative for the next several weeks. The 14-point enterprise preference for non-Nvidia chips will draw responses from AMD and Intel, and Nvidia’s $99 billion investment framing is likely to intensify. Security teams and AI coding assistant vendors will also be working through the GitLost disclosure and its patch timeline. It was a dense week, and the open threads heading into the next one are substantial.
