Anthropic Bets on Consulting, Not Just Models, With $1.5B Wall Street Venture

The new firm—not yet named—draws approximately $300 million each from Anthropic, Blackstone, and Hellman & Friedman, with Goldman Sachs contributing around $150 million. A broader consortium including Apollo, General Atlantic, GIC, Leonard Green, and Sequoia provides both capital and a built-in pipeline of portfolio companies. The structure mirrors Palantir's forward-deployment model rather than conventional SaaS. As Blackstone COO Jon Gray put it, the venture aims to address 'one of the most significant bottlenecks to enterprise AI adoption'—the scarcity of engineers who can implement frontier AI systems quickly. Anthropic CFO Krishna Rao noted that 'enterprise demand for Claude is significantly outpacing any single delivery model.' The timing is notable: hours before Anthropic's announcement, Bloomberg reported OpenAI was raising funds for a similar venture called The Development Company. Both labs appear to have concluded that model capability is no longer the constraint—operational integration is. For every dollar companies spend on software, they spend six on services, and AI labs now want a share of that multitrillion-dollar market.

The seventh edition of Stanford HAI's annual report tracks AI across technical, economic, and societal dimensions. The headline finding on agents is dramatic—from barely functional to approaching human-level performance on real computer tasks involving file manipulation, application navigation, and multi-step workflows. The report also captures what researchers call the 'jagged frontier' of AI capabilities: Gemini Deep Think earned a gold medal at the International Mathematical Olympiad, yet the best model reads analog clocks correctly only 50.1% of the time. On coding, SWE-bench Verified performance rose from 60% to near 100% of the human baseline in a single year. The economic chapter reveals a disconnect between capability and deployment. Generative AI is now used in at least one business function at 70% of surveyed organizations, but AI agent deployment remains in single digits across nearly all functions. The report notes that heavy AI reliance may carry 'long-term learning penalties that slow skill development over time.' Industry produced over 90% of notable frontier models in 2025, while the Foundation Model Transparency Index saw average scores drop from 58 to 40—the most capable models disclose the least.

The research, published in early April, used the Common Crawl archive to identify prompt injection patterns planted in publicly accessible HTML. Google's team categorized the attempts into pranks, SEO manipulation, helpful guidance, deterrence, and explicitly malicious payloads. Most attacks remain unsophisticated—instructions like 'ignore previous instructions' embedded in page source—but the upward trend is clear. Google warned that 'this could change soon' as AI systems become more capable and threat actors automate operations with agentic AI. The disclosure coincided with a Forcepoint report documenting IPI attempts aimed at financial fraud, including payloads that embed fully specified PayPal transactions and step-by-step instructions for AI agents with payment capabilities. As one analyst put it: 'A browser AI that can only summarize is low-risk. An agentic AI that can send emails, execute terminal commands or process payments becomes a high-impact target.' Google has detailed its layered defense strategy for Gemini, including markdown sanitization, suspicious URL redaction, and a user confirmation framework for sensitive actions. But the fundamental tradeoff persists: defenses that reduce attack success also degrade model utility.

The Pentagon's announcement covers Impact Level 6 and Impact Level 7 network environments—the highest classification tiers—and aims to 'streamline data synthesis, elevate situational understanding, and augment warfighter decision-making.' Pentagon CTO Emil Michael explicitly referenced Anthropic when explaining the multi-provider strategy: 'We learned that that one partner didn't really want to work with us in the way we wanted to work with them.' Anthropic's Claude was previously the only frontier AI model deployed on classified Pentagon networks through Palantir's Maven toolkit. The relationship soured after Anthropic refused terms that would allow military use for 'all lawful purposes,' including autonomous weapons. The Trump administration subsequently labeled Anthropic a 'supply chain risk'—a designation typically reserved for companies linked to foreign adversaries. Anthropic sued and won a preliminary injunction in federal court. More recently, CEO Dario Amodei met with White House Chief of Staff Susie Wiles after Anthropic unveiled its Mythos cybersecurity tool. The NSA is reportedly using Mythos despite the broader ban, suggesting the standoff may be softening. For Anthropic, the exclusion represents both substantial lost revenue and a test of whether safety-first positioning is sustainable against government pressure.