Anthropic Emerges Ahead in AI Valuation Race

人工智能巨头Anthropic提交了在美国上市的保密申请

Anthropic has confidentially filed for an initial public offering, signaling the AI startup’s move toward a public market debut. The company, known for developing large language models such as Claude, joins a wave of AI firms seeking to monetize and scale amid high investor interest in generative AI. A confidential filing lets Anthropic begin the IPO process with the SEC while keeping details private until a public filing; the move could raise capital for model development, infrastructure, and commercial expansion, and offer liquidity to investors. The decision matters for competition among AI cloud and model providers and may influence valuations and hiring across the sector.

Mike Isaac / New York Times : Anthropic says it has confidentially filed for an IPO, joining OpenAI and SpaceX in preparing to go public this year — The artificial intelligence company has seen explosive growth over the last year thanks largely to new technology that can automatically write computer code.

Anthropic : Anthropic says it has confidentially filed for an IPO with the SEC — Today, Anthropic, PBC confidentially submitted a draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission for a proposed initial public offering of our common stock.

Anthropic, the AI startup focused on large language models, confidentially submitted a draft Form S-1 to the U.S. SEC as a step toward a potential initial public offering. The filing gives Anthropic the option to go public once the SEC completes its review, though share count, pricing and timing remain undecided and depend on market conditions. The company framed the announcement under Rule 135 and emphasized this is not an offer or solicitation. This move matters because an IPO would broaden public access to investment in a leading AI company, influence capital flows and competitive dynamics among generative-AI firms, and signal market appetite for AI platform providers.

Anthropic disclosed that its browser-based agent was hijacked in 31.5% of red-team prompt-injection attempts against Opus 4.8 before safeguards engaged, a headline figure that contrasts with much lower rates when safeguards are active (0.5%) or thinking is off (0%). Anthropic published a 244-page system card breaking results out by four agentic surfaces and showing wide variance: coding environments saw single-attempt success as low as 7.03% (2.09% with safeguards), while web/browser contexts were far more vulnerable. OpenAI, Google and Meta provided inconsistent, noncomparable disclosures: OpenAI reported a single robustness score for connectors (GPT-5.5 = 0.963), Google placed details in a separate safety framework, and Meta offered no closed-model card. The lack of industry standards for measuring prompt injection complicates procurement and risk management for enterprises and defenders.

Anthropic has overtaken OpenAI to become the most valuable AI startup, sparking broad discussion among developers about model quality versus marketing and user experience. Hacker News commenters debated whether measurable performance differences or branding and UX (meetups, integrations like VS Code) drive adoption, with users citing models such as Claude, GPT-5.5 and Opus variants in real-world coding workflows. Commenters argued that developer preferences hinge on factors beyond raw output—interaction style, need for steering, tooling, speed, cost and multi-model switching—while noting benchmarks still matter for specific tasks like optimization. The shift matters because valuation reflects market confidence and influences hiring, investment, and competitive dynamics across AI platforms and developer tooling.

OpenAI and Anthropic’s push with powerful base models has many founders asking whether there’s any room left to build AI applications. Joe Schmidt argues that while the labs will dominate the “Yellow Brick Road” — horizontal, capability-driven apps like code generation or generic agents — meaningful startup opportunity remains in vertical, domain-specific software that requires trust, compliance, operational scaffolding, and deep integrations. The labs are investing in enterprise customizations and joint ventures, signaling they don’t expect models alone to solve every use case. Startups should avoid simple connector-plus-agent plays that the labs can replicate, and instead focus on specialized workflows, regulatory needs, and productized integrations where differentiation and customer ownership are defensible.

Anthropic’s public reporting uses a specific “run-rate revenue” metric quoted by Karen Kwok for Reuters Breakingviews: it annualizes recent consumption sales by taking the last 28 days of usage-based customer charges and multiplying by 13, then adds annualized subscription revenue (monthly subscription take multiplied by 12). The formula, attributed to an unnamed source, clarifies how the AI startup aggregates consumption and subscription streams to present a forward-looking revenue run rate. This matters for investors, competitors and analysts assessing Anthropic’s growth, unit economics and product-market fit relative to rivals like OpenAI as usage-based AI pricing becomes central to commercial models. The quote was posted on Simon Willison’s weblog on May 31, 2026.

Anthropic published detailed documentation of the sandboxing techniques used across its Claude products, and Simon Willison highlights the overview as a rare, useful disclosure. The company uses gVisor for Claude.ai, Seatbelt on macOS and Bubblewrap on Linux for Claude Code (local), and full VMs via Apple Virtualization and HCS on Windows for Claude Cowork. Controls include process sandboxes, VMs, filesystem boundaries, and egress restrictions to prevent credential exfiltration and other attacks; Willison notes past missed risks like the api.anthropic.com/v1/files exfiltration vector. The post also mentions Anthropic’s open-source Anthropic Sandbox Runtime (srt) as mature enough for further evaluation. This level of transparency matters for assessing trust and security in AI platforms.

Anthropic has reportedly surpassed OpenAI in valuation after a fresh funding round, marking a shift in the AI startup landscape. Investors have pushed Anthropic’s worth above OpenAI’s by backing its safety-focused, large-language-model strategy and commercial partnerships. This matters because it signals investor confidence in alternatives to OpenAI’s models and could reshape competitive dynamics among foundation-model developers, enterprise AI sellers, and cloud partners. Key players include Anthropic and OpenAI; the development highlights investor appetite for safety-aligned architectures and diversified product offerings. The change may influence hiring, model licensing, and platform integrations across the AI ecosystem as enterprises and cloud providers reassess supplier choices.

Anthropic的市值现已超过OpenAI - Gizmodo

A report says an unnamed, very large company accidentally spent $500 million in a single month on Anthropic’s Claude AI after failing to set usage limits on employee Claude licenses. Axios cites an AI consultant and ties the incident to broader corporate worries about runaway AI costs, token-heavy agent tools, and employees gaming internal metrics by overusing models. The episode follows other billing mishaps — a Google Cloud customer hit with an $18,000 bill after a breach and an OpenClaw developer burning $1.3M in OpenAI tokens — and has prompted firms like Amazon to remove AI usage leaderboards. The case highlights governance, cost controls, and ROI challenges as enterprises scale AI deployments.

Anthropic’s public definition of “run-rate revenue” was quoted by Karen Kwok for Reuters Breakingviews and reposted by Simon Willison on May 31, 2026. According to the quote, Anthropic calculates run-rate revenue by annualizing the last 28 days of consumption-based sales (multiply by 13) and adding annualized monthly subscription revenue (multiply by 12). The detail matters for how the AI company presents near-term revenue momentum to investors and the market, because mixing short-term consumption extrapolation with subscription annualization can influence perceived growth and valuation. The note is a sourced quotation, attributed to “a person familiar with the matter.”

A report by Axios, highlighted by Tom's Hardware, says an unnamed large company accidentally spent $500 million in a single month on Anthropic’s Claude AI after failing to set usage limits on employee Claude licenses. The episode illustrates a growing corporate pain point: runaway AI consumption and weak governance driving massive cloud/AI bills without clear return on investment. The article ties the incident to broader trends—employee “tokenmaxxing,” costly agentic AI workflows, and other billing fiascos such as a Google Cloud $18,000 surprise and a $1.3M OpenAI token burn—prompting firms like Amazon to remove internal AI leaderboards. The story matters because it underscores urgent needs for cost controls, monitoring, and policy around enterprise AI usage.

Anthropic 为何估值超过 OpenAI？答案就在这一个词中 - inc.com

Anthropic的估值已超越OpenAI。但有一点需要了解。 - Investor's Business Daily