AI model costs skyrocket, raising questions about access and control

• Anthropic’s new Claude models cost double previous versions, with a two-tier system that restricts high-risk queries
• Industry data shows AI companies like OpenAI and Anthropic are losing billions annually despite rising prices
• Experts predict top-tier AI tools will cost $20,000+ per user per year within two years
• Cursor, a major AI coding app, sends 100% of its revenue to Anthropic and remains deeply unprofitable
• Local AI models and open-source alternatives emerge as counterpoints to increasingly expensive and controlled systems

The price shock heard round the tech world

The era of affordable, unrestricted artificial intelligence may be ending. On June 22, 2026, Anthropic released Claude Fable 5 for commercial users and Mythos 5 for restricted access, charging $10 per million input tokens and $50 per million output tokens—roughly double the cost of Claude Opus 4.8. The pricing represents a fundamental shift in how AI companies are monetizing their most advanced systems, raising questions about who will have access to cutting-edge AI capabilities.

Industry analysts and technology observers are calling the move a “rug pull,” arguing that users are paying premium prices while being silently downgraded to older models on queries deemed high-risk. Anthropic has stated that fewer than 5% of sessions trigger this downgrade, but critics fear the percentage will grow.

The exponential cost curve

AI coding tool subscription prices are following an exponential trajectory, according to data compiled by industry observers. The cheapest usable tier of Claude Code now costs $100 monthly, up from the $10 per month that GitHub Copilot charged just years ago. OpenAI has reportedly discussed charging $20,000 monthly for PhD-level research agents with investors.

The cost increases reflect a harsh economic reality: every major generative AI company is losing money. OpenAI lost $5 billion in 2024, while Anthropic lost $5.3 billion. Even when removing training costs from OpenAI’s 2024 revenue, the company would still have lost $2.2 billion. Perplexity AI spent 164% of its revenue on cloud computing services in 2024.

The two-tier system debate

Anthropic’s split between Fable 5 and Mythos 5 has drawn particular scrutiny. Mythos 5, described as a highly vetted, government-coordinated tier, is open to approximately 200,000 defense organizations and critical-infrastructure operators. Sources claim safety restrictions do not apply to this elite tier. Fable 5 is accessible to anyone via cloud API, Amazon Bedrock, Google Cloud and Pro/Max/Team subscription tiers.

Critics argue the distinction represents a privacy concern, with Anthropic moving toward identity-verifying users based on their queries. This “shotgun KYC” approach, as some have termed it, ties directly to the company’s upcoming IPO and pressure from investors who cannot risk liability from bad actors using the platform.

From free speech to controlled access

The current situation echoes earlier shifts in internet governance. At Google, the transition from a free speech platform to one tightly controlled by certain narratives created a distinction between “authoritative content”—information from sources like the BBC and Wikipedia—and everything else. OpenAI is now feeding its models vast amounts of this so-called authoritative data.

As algorithms grow larger and are exposed to more diverse information, they begin recognizing inconsistencies in their training data. The goal is not just to replicate the World Wide Web but to compress it into a manageable format that generates coherent text. However, contradictions pose a significant challenge because they cannot be accepted as truth.

This dynamic mirrors themes from George Orwell’s “1984,” where governments manipulate information to maintain control. In the current environment, AI companies are filtering out contradictory data sources, potentially limiting access to alternative perspectives on topics like nutrition, herbs, permaculture and health practices outside conventional medical recommendations.

The local alternative emerges

Technology advancements offer a counterpoint to centralized AI control. Executable files for AI models can now run locally on personal computers through platforms like torrents or decentralized networks such as BitTorrent. Smaller-scale yet powerful multimodal language models—those with 13 billion parameters—can operate on individual devices.

The cost of fine-tuning and training these models is decreasing rapidly. Industry observers predict that within two years, the cost of building and distributing large language models could drop to around $20,000 and continue decreasing further. Once scientists uncover how human brains efficiently perform backpropagation for learning reinforcement, this knowledge could be replicated in digital systems, drastically reducing the cost, energy consumption and time required to train neural networks.

A fork in the digital road

The AI industry stands at a critical juncture. On one path lies increasingly expensive, controlled systems that serve elite interests and limit access to contested information. On the other lies decentralized, local AI models that anyone can deploy and customize. The economic realities of generative AI—with every major provider losing billions—suggest that current pricing models are unsustainable. Whether the industry pivots toward mass-market affordability or doubles down on premium, controlled access will determine not just the future of AI but the nature of information access itself.

Sources for this article include:

TheDailyBell.com

DanielPaleka.com

WheresYourEd.at