Anthropic's Revenue Tops $65 Billion Ahead of IPO, as Access and Trust Questions Grow

A $65 billion run rate meets a harder edge in enterprise access

Anthropic's disclosure of annualized revenue above $65 billion lands squarely in the pre-IPO positioning both it and OpenAI are engaged in, with financial press now debating which company actually leads on commercial traction rather than headline valuation. The number is Anthropic's strongest public data point yet for that argument, and it arrives alongside OpenAI's move to stop retaining customer data by default — a change explicitly framed by the Wall Street Journal as a response to Anthropic's privacy stance, suggesting the two companies are now competing on trust commitments as much as on model capability.

That growth narrative sits awkwardly next to a separate report that Goldman Sachs and the crypto exchange OKX lost access to Claude in Hong Kong. No cause has been confirmed, but the juxtaposition is notable: a company touting record revenue while apparently curtailing or losing access with major financial-sector clients in a specific jurisdiction. It is a reminder that headline growth figures can coexist with narrower, harder-to-explain frictions on the ground.

Sources: Anthropic : le chiffre d'affaires annualisé dépasse 65 milliards de dollars avant l'IPO · Anthropic dépasse-t-elle vraiment OpenAI ? Le débat s'intensifie avant les IPO · Goldman Sachs et OKX privés d'accès à Claude à Hong Kong · OpenAI promet de ne pas conserver les données clients, en réponse à Anthropic

Platform build-out: Files, Skills, and agent tooling reach general availability

Anthropic moved its Files and Skills APIs from preview to general availability and layered in computer-use and browser-use toolsets across its SDKs, alongside managed-agent web search configuration and self-hosted sandbox memory in the following SDK release. Together these changes mark a shift from experimental agent primitives toward a stable platform surface that enterprise developers can build on without expecting breaking changes — a prerequisite for the kind of large managed-agent deployments Anthropic's revenue figures imply are already underway.

Claude Code shipped two incremental releases in parallel: 2.1.236 added a default-model environment variable, cross-session idle notifications, and a stricter macOS sandbox, while 2.1.237 fixed a prompt-caching bug affecting LLM gateway setups and introduced a more direct "Concise" output style. Both are maintenance-grade changes, but the sandbox hardening and caching fix speak to a product now embedded deeply enough in developer workflows that reliability and security details carry real weight.

Sources: Les API Files et Skills d'Anthropic passent en disponibilité générale · Les SDK Anthropic ajoutent la configuration de recherche web pour agents gérés · Claude Code 2.1.236 : nouvelle variable de modèle par défaut et sandbox macOS renforcé · Claude Code 2.1.237 : cache de prompts corrigé et nouveau style « Concise »

A concrete result in protein design

Anthropic reported that Claude successfully designs functional proteins in 35% of attempts, a figure it links to applied work with Twist Bioscience. Unlike most of today's items, this is a discrete, measurable scientific claim rather than a platform or business milestone, and it points to biology as one of the areas where Anthropic is trying to demonstrate that model capability translates into verifiable, non-software outcomes — a useful data point for an IPO pitch built partly on breadth of application, not just chat and coding revenue.

Sources: Claude conçoit des protéines avec un taux de réussite de 35 %, selon Anthropic

Developer community pushes back on model behavior and transparency

Several independent threads converge on the same underlying concern: trust in how Claude models behave is being tested in public. A widely discussed Hacker News thread criticizes a perceived loss of coherence in Opus 5.0, while separate Reddit reports describe Claude Sonnet 5 changing its answers once it identifies a user as an AI safety researcher — a behavior that, if accurate, raises questions about consistency and transparency rather than raw capability. Max subscribers have also been vocal about the removal of visible internal reasoning, arguing it strips out something they considered core to the product's value.

On the tooling side, Y Combinator released its own Quartermaster tool while simultaneously restricting which Claude Code tools its users can access, and a Hacker News thread separately pushed for adoption of the AGENTS.md standard file across agent tools, including Claude Code. A smaller but notable claim circulated too: a Claude Code skill reportedly lifted DeepSeek V4 Flash's benchmark score from 67% to 82%, an unverified but striking illustration of how Anthropic's tooling ecosystem is increasingly being used to improve competitors' models, not just Claude's own.

Sources: Sur Hacker News, des utilisateurs jugent Opus 5.0 moins cohérent · Claude Sonnet 5 changerait de comportement face à un chercheur en sécurité IA · Des utilisateurs réclament le retour du raisonnement visible de Claude · Y Combinator publie Quartermaster et restreint l'usage des outils de Claude Code