Two things happened in the past week that every business building on Anthropic’s Claude should know about. Taken individually, they look like operational hiccups. Together, they point to something builders need to factor into their platform decisions.

What Happened

On April 21, Anthropic quietly removed Claude Code from its $20-per-month Pro subscription plan. No announcement. No email to affected users. No changelog entry. Developers noticed when they compared archived support documentation to the updated version.

The implication: Pro users who relied on Claude Code for their work would need to upgrade to the Max 5x tier at $100 per month — a five-fold price increase — to keep their existing workflow running.

Social media responded immediately. One post flagging the change accumulated nearly 900,000 views within hours. By April 22, Anthropic reversed the decision. Amol Avasare, Anthropic’s Head of Growth, confirmed the change was a test affecting roughly two percent of new users.

A two percent test with zero communication, until the internet noticed.

The Tokenizer Story

Separate from the Pro plan incident, Anthropic buried a footnote in the Claude Opus 4.7 launch announcement: the new model uses a new tokenizer that “can consume up to 35% more tokens for the same text.”

The list price for Opus 4.7 is unchanged from its predecessor — $5 per million input tokens, $25 per million output tokens. But if the same prompt consumes 35% more tokens, the effective cost of running the same workload is 35% higher. Developer reports published after launch put the real-world increase somewhere between 32 and 47 percent for typical tasks.

This is not a new practice in the industry. OpenAI, Google, and others have all made architectural changes that affected token consumption. But the disclosure came in a footnote, not in the headline of the announcement. For teams running Claude in production at meaningful scale, this is a billing change that arrived without emphasis.

The Pattern

Anthropic now generates more than $30 billion in annualised run-rate revenue, up from roughly $9 billion at the end of 2025. That is extraordinary growth. It also means the company faces a familiar scaling pressure: monetising heavy users more aggressively while keeping entry-level access attractive enough to grow the user base.

The Pro plan incident and the tokenizer change are both visible expressions of that pressure. The reversal on Claude Code access was the right call — but the initial decision and the absence of communication was not a confidence-building move. The tokenizer disclosure was technically accurate; it was the placement that was the problem.

Combined with Anthropic’s decision in early April to block third-party agentic tools from Claude Pro and Max subscriptions — a move that pushed some power users from $20 plans to costs 50 times higher overnight — the past month has seen a pattern of changes that primarily affect the people most invested in building on Claude.

What This Means for Business

If you are currently building workflows, internal tools, or products on top of Claude, a few things are worth working through now rather than later.

Understand your actual cost structure. If you upgraded to Opus 4.7 without adjusting for the tokenizer change, your costs have moved. Pull the last 30 days of token usage and compare it against equivalent prompts on the prior model. The delta will tell you whether the 35% ceiling matters for your use case.

Build for portability. The OpenAI-Microsoft exclusivity ending this week, Google’s growing Gemini 3.1 Ultra capabilities, and AWS’s push to distribute models across all major providers means the multi-cloud AI market is increasingly real. Building in hard dependencies on any single provider is a strategic risk worth managing.

Treat AI pricing as a live variable. AI infrastructure is still being priced in real time by companies under enormous growth pressure. Budget assumptions made six months ago are probably wrong. Building monthly cost reviews into your AI program is now standard operating practice, not optional.

Watch the communication quality. Anthropic builds some of the most capable models available and has a strong safety track record. But how a company communicates pricing changes is a signal about how it views its customers. The past month has been noisier than it needed to be. That is worth watching over the next quarter.

The Bigger Picture

None of this suggests abandoning Claude — Opus 4.7 remains one of the best models for complex reasoning and coding tasks, and the Claude Code Routines feature launched earlier this month is genuinely useful for teams running autonomous workflows.

But it does reinforce something worth repeating: picking an AI model is not the same as picking an AI strategy. The tools will keep changing. The costs will keep evolving. The work of integrating AI into a business involves understanding the commercial dynamics of the providers you depend on, not just the technical capabilities.

Right now, the commercial dynamics at Anthropic are moving fast. Stay close to your usage data and keep your architecture as flexible as the technology allows.