AI subscriptions are becoming less attractive

하이라이트

• Anthropic is testing the removal of Claude Code access from Pro plans, requiring more expensive subscription tiers.

• GitHub stopped new signups for Pro, Pro Plus, and Student plans while tightening usage limits due to heavy consumption.

• Claude Opus 4.7 API pricing is $5 per million input tokens and $25 per million output tokens.

• GPT-5.4 API pricing is $2.50 per million input tokens and $22.50 per million output tokens.

• Agentic workflows create significantly higher compute demands than occasional chat sessions, as they process millions of tokens rapidly.

• OpenAI raised $122 billion, yet projections suggest this only provides approximately 18 months of operational runway without sustained profitability.

• AI compute scarcity, specifically for high-demand networking gear, memory, and energy, is limiting the available capacity for training and inference.

타임라인

1. 00:00 - 02:17

   Shift in Subscription Access and Limits

   Evidence indicates a clear trend toward tightening constraints on AI subscriptions. Recent actions by Anthropic and GitHub demonstrate a move to protect compute resources as users shift toward more intensive usage patterns. These restrictions follow a pattern of stricter limits and perceived performance degradation across major AI services.

   • Anthropic is testing the restriction of Claude Code to higher-tier plans for 2% of new prosumer signups.
   • GitHub paused new signups for Copilot Pro and Student plans to implement stricter individual usage limits.
   • Companies are aggressively curtailing usage outside of official platforms to protect subscription margins.
2. 02:17 - 09:44

   Economics of Token Consumption

   The current economic model for AI subscriptions is unsustainable when users transition from occasional chat to high-volume agentic workflows. API pricing from providers like Anthropic and OpenAI highlights the massive disparity between flat-rate subscriptions and the actual cost of high-volume inference. Companies are currently subsidizing heavy users, a practice that cannot continue long-term without compromising business viability.

   • Subscription models rely on the majority of users not exhausting their allotted usage limits.
   • API pricing serves as a benchmark for the true cost of inference, with Claude Opus 4.7 costing $25 per million output tokens.
   • Training costs are one-time capital expenditures, while inference costs are ongoing and directly linked to usage volume.
3. 09:44 - 13:58

   Compute Scarcity and Infrastructure Pressures

   Infrastructure requirements have become a critical bottleneck for AI companies. The need for massive GPU clusters creates extreme pressure on memory, networking, and electricity supply. Because energy cannot be easily scaled via the public grid, building new capacity is a slow and capital-intensive process that limits the overall available compute for both training and inference.

   • High demand for specialized networking gear, memory, and energy is driving up the cost of AI infrastructure.
   • Data center expansion is constrained by power grid limitations, forcing reliance on expensive, off-grid energy solutions.
   • Providers must now split scarce compute capacity between model training and customer-facing inference.
4. 13:58 - 18:03

   Future Outlook for AI Costs

   The transition to agentic workflows has fundamentally altered the economics of AI usage. Models now generate far more tokens per task due to internal reasoning processes, vastly increasing the cost of inference. As companies move past their initial cash runways, they will likely segment subscriptions, offering lower-cost options for casual users while significantly raising prices for professional, agent-heavy workloads.

   • Agentic workflows and 'thinking' models require significantly more tokens than simple chat interfaces.
   • Enterprise-focused companies like Anthropic are feeling usage-related cost pressures earlier due to their customer base.
   • Subscription prices for professional agentic tools will likely increase to compete with the cost of human labor.