Anthropic Widens Security Footprint as Enterprise Clients Question AI Costs Ahead of IPO

Anthropic files for IPO amid enterprise cost scrutiny. Glasswing cybersecurity scales to 150 partners, but Microsoft cancels licenses. New agent billing credits begin June 15.

Just as Anthropic pushes deeper into critical infrastructure with its Project Glasswing expansion, a countercurrent is building among the corporate customers it most needs to impress. The company officially kicked off its initial public offering process on June 1, 2026, filing a confidential registration draft with the SEC, but the narrative for public investors is far from settled: can Anthropic convince the market that enterprise demand will keep surging, or will mounting cost scrutiny cap its growth?

The IPO timeline remains unannounced, pending SEC review and market conditions. Price and share count have not been set. But in the interim, Anthropic is working to shape a story that goes beyond AI hype. The Glasswing cybersecurity program, initially launched with around 50 partners, is now being extended to 150 partners across more than 15 countries. New participants include operators in electricity, water, healthcare, communications, and hardware — sectors whose codebases underpin systems that serve more than 100 million people. Apple, Nvidia, Microsoft, CrowdStrike, and Palo Alto Networks were already on board; Rubrik has confirmed it is among the new additions.

Access is not automatic. Anthropic requires partners to meet its own security criteria before they can use Claude Mythos to scan for vulnerabilities. The first cohort of roughly 50 partners identified more than ten thousand high- or critical-severity flaws — a figure that Anthropic is using as a sales differentiator. The message is deliberate: Claude is not a lab experiment but a production-grade tool for complex enterprise environments, precisely the budgets that justify a high IPO valuation.

Yet even as Anthropic deepens its ties to critical infrastructure, it faces pushback in its core enterprise market. Microsoft, one of the largest corporate users of Claude licenses, has canceled a significant portion of them, citing standardization. Uber’s management has publicly stated that AI spending is increasingly hard to justify. A Bain & Company survey of 951 companies worldwide found that nearly 40% of those measuring cost savings from AI reported less than 10% improvement, well short of the 11–20% many had targeted. Still, 90% of those companies plan to increase their AI budgets — a sign that growth may continue, but at a more measured pace.

To address these pressures, Anthropic is restructuring how it bills for agent usage. Starting June 15, 2026, Claude Agent SDK usage and the claude -p command will no longer count against regular usage limits. Instead, eligible users receive a separate monthly credit: $20 for Pro and Team Standard users, $200 for Max-20x and Enterprise Premium users. The credits apply to Agent SDK, GitHub Actions, and third-party apps, but they expire at month-end and cannot be transferred. The move effectively separates interactive usage from programmatic agent workloads, giving Anthropic more control over scaling costs.

The API pricing itself underscores the potential for cost escalation. Claude Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens; Claude Opus 4.8 to 4.5 ranges from $5 input to $25 output. Autonomous coding or analysis workflows can rapidly generate high token volumes, which is both Anthropic’s revenue engine and its vulnerability. If customers shift simple tasks to cheaper models or ration usage, the growth trajectory could flatten.

Anthropic’s financials reflect the high-stakes balancing act. Annualized revenue surpassed $4.7 billion in May 2026, up from roughly $10 billion the prior year — a dramatic acceleration that demands a correspondingly large addressable market. The Series H round raised $6.5 billion at a post-money valuation of $61.5 billion. But data from EntelligenceAI indicates that only 18% of spending on advanced AI coding tokens translates into shipped products, raising questions about the clarity of return on investment.

Infrastructure costs add another layer. Anthropic is paying SpaceX $1.25 billion per month through May 2029 for access to more than 300 megawatts of compute power and over 220,000 Nvidia GPUs, including next-generation GB200 hardware in the Colossus 2 expansion. That spending explains the rapid product rollout — Claude Security entered public beta on April 30, 2026, scanning codebases with Claude Opus 4.7, and a compliance interface linking to 28 platforms including CrowdStrike, Okta, and Zscaler followed in late May.

The competitive landscape is equally intense. OpenAI is building government ties through its OpenAI for Countries initiative, with the Stargate UAE project as a one-gigawatt cluster in Abu Dhabi involving Oracle, Nvidia, and SoftBank. Anthropic’s narrower cybersecurity focus via Glasswing overlaps with those same government digital and security budgets, but the race for sovereign AI contracts is far from won.

For established vulnerability-scan vendors, the Glasswing expansion raises the stakes. After Claude Code Security was unveiled in February, cybersecurity ETFs and standalone security stocks came under pressure as the market priced in the risk that AI could displace traditional scanning and compliance workflows. With Anthropic now connecting Claude directly to security platforms and generating patches, the value chain is shifting. Providers without a strong AI layer could face margin compression.

Anthropic’s IPO will ultimately be judged on whether customers see Claude spending as an investment or a cost line to be cut. That question will weigh more heavily on the opening valuation than any growth rate. For now, the company is using the IPO runway to demonstrate scale, security credentials, and global reach — and Project Glasswing is a key piece of that narrative, not merely a product update.