AI Subscription Models Under Strain: Anthropic and GitHub Signal End of Heavily Subsidized Usage

Anthropic recently initiated a limited test for 2% of new prosumer signups, restricting access to Claude Code within its Pro plan and directing users towards more expensive tiers. This move, despite Anthropic’s explanation as a “small test,” quickly drew community criticism for aligning with a perceived trend of diminishing subscription value, characterized by stricter usage limits and concerns over degrading model performance. This sentiment is amplified by Anthropic’s earlier actions to curb external usage of its subscriptions and further reinforced by GitHub’s recent announcement. GitHub Copilot Pro, Pro+, and Student plans now face tightened usage limits, with Opus models no longer available in the Pro tier, signaling a broader industry shift away from the previously held promise of heavily subsidized, “unlimited” AI usage.

The underlying economic reality for AI providers like Anthropic and OpenAI is becoming increasingly apparent: the days of unlimited, low-cost AI inference are unsustainable. Subscription models, like any utility, are profitable when the majority of users do not fully exploit their allowances. However, the rise of “agentic workflows”—long-running, parallelized coding sessions—dramatically increases token consumption, often burning millions of tokens quickly, far beyond typical chat sessions. This, coupled with global compute scarcity driving up costs for memory, networking, energy, and data center infrastructure, puts immense pressure on providers. API pricing pages reveal the true cost of inference (e.g., Claude Opus at $5/M input, $25/M output tokens), making current subscription prices, which offer many millions of tokens for a fraction of that cost, financially challenging. To maintain viability and a competitive edge, providers must now balance aggressive market share capture with the need to cover significant training and ongoing inference expenses, indicating a future with stricter limits, higher prices, and potentially differentiated subscription tiers tailored for specific, resource-intensive use cases.