Anthropic is Preparing for IPO and We Should Be Worried

Anthropic is starting to act like a company preparing for public markets: protecting margins, tightening access. The gap between the brand and the behavior is now too obvious to ignore.

Claude Code

Starting June 15, 2026, Anthropic says Claude Agent SDK and claude -p usage will stop counting toward normal Claude subscription limits. It will draw from a separate monthly Agent SDK credit instead. After that credit runs out, additional usage moves to usage credits at standard API-style rates if the user has enabled them. Anthropic’s help center says the new credit covers the Agent SDK, claude -p, Claude Code GitHub Actions, and third-party apps built on the Agent SDK, while interactive Claude Code stays in the normal subscription bucket.

That difference matters because claude -p is not some unusual enterprise edge case. It is the normal CLI workflow: you give the tool a prompt and get output back. Developers use that form in shell scripts, review workflows, automation, CI, and small one-off tasks where opening a full interactive session makes no sense.

Third-party workflows

In April, Anthropic cut off Claude Pro and Max subscription access for third-party agent frameworks such as OpenClaw because those tools strained its compute and engineering resources. Users could still use Claude through API-style billing, but the flat-rate subscription path was no longer available for that kind of workflow.

OpenCode, Pi, and Claude access

The issue is broader than OpenClaw. The Register reported that developers had been using third-party harnesses such as OpenCode and Pi to interact with Claude Code through their Claude subscriptions. The practical result is simple: your Claude subscription works in Anthropic’s own surfaces, Claude.ai, Claude Code, and Cowork. Use OpenCode, Pi, OpenClaw, or another harness, and that subscription no longer counts. You need extra usage or an API key. The plan starts to look like a pass for Anthropic’s preferred interfaces, not general Claude access.

Platform

Anthropic’s commercial terms prohibit customers from using the service to build competing products, train competing AI models, resell the services without approval, or reverse engineer the services. That is a defensible business choice. No company wants to help fund its competitors.

But Anthropic has also shown that it will enforce those limits. Wired reported in 2025 that Anthropic revoked OpenAI’s API access, and that it had restricted Windsurf’s direct access after rumors that OpenAI might acquire it.

Anthropic talks a lot about safety and responsible deployment, but it has not released a major open-weight Claude model.

Mythos, Anthropic, and controlled access

The company introduced Claude Mythos Preview through Project Glasswing as a cyber model serious enough to keep behind a partner program. Anthropic may see that as consistent with its risk model. But the public evidence so far does not show Mythos living up to that. AISI later said GPT-5.5 reached a similar level on its cyber evaluations and scored slightly higher on expert tasks, 71.4% versus 68.6%. Daniel Stenberg’s curl test also cut through the aura: Mythos found one low-severity curl issue, and he did not see evidence that it found bugs at a level other AI analyzers had not already reached. “Too dangerous for broad access” also justifies Anthropic’s preferred setup: selected partners, enterprise and government channels, and tight control.

Code quality

In April, Anthropic published a postmortem on Claude Code quality reports. The company said several product-layer changes had caused user reports of lower quality. One change lowered Claude Code’s default reasoning effort from high to medium. Anthropic later called that the wrong tradeoff and reverted it. Other issues involved clearing older thinking from sessions and a prompt change that hurt coding quality.

Public image

Anthropic still presents itself as the careful AI company. In February, it promised that Claude would remain ad-free, casting it as a place to think instead of another ad market.

Anthropic is also expanding into enterprise, government, and defense channels. In 2025, the company announced a $200 million Department of Defense agreement and said Claude was already integrated into mission workflows with partners like Palantir, supporting U.S. defense and intelligence organizations on classified networks.

That does not make Anthropic wrong by itself. Large AI companies will sell to governments, enterprises, and defense-adjacent customers. But it does make the brand harder to take at face value. Anthropic increasingly wants the reputation of a careful public-interest lab while acting like a normal, ambitious platform company.

Anthropic recently announced a SpaceX partnership giving it access to all compute capacity at SpaceX’s Colossus 1 data center: more than 300 megawatts and over 220,000 NVIDIA GPUs within the month. It also listed major compute commitments with Amazon, Google and Broadcom, Microsoft and NVIDIA, and Fluidstack.

That scale changes the business. As compute rises, Anthropic will feel more pressure to monetize usage correctly, tighten workflows that were previously underpriced, and satisfy investors with enterprise contracts, predictable revenue, and platform control.

IPO

Anthropic is becoming a public-market platform company.

This is why Anthropic is making flat subscriptions less useful for automation, splitting CLI and agent usage into new billing pools, raising the cost of third-party workflows, restricting strategic access, keeping Claude closed, expanding into government and defense, and buying enormous compute capacity.

Each decision can be defended on its own. Together, they describe a company becoming more closed, more controlled, and more aggressive about monetization.