The AI Price Divide: Who Gets the Best Models?

As frontier AI gets more expensive, not cheaper, a new inequality emerges

Every technology revolution was supposed to get cheaper over time. Computers shrank from room-sized machines to pocket devices. Internet access went from expensive luxury to basic utility. Cloud storage dropped from dollars per gigabyte to cents.

AI is different.

The newest frontier models are getting more expensive, not cheaper. Anthropic's leaked Claude Mythos (code-named "Capybara") sits above Opus in a new pricing tier. OpenAI's o1-pro costs $150 per million input tokens — ten times Claude Opus. The best AI is becoming a luxury product.

This creates a divide that didn't exist in previous technology revolutions: the rich get better intelligence, while everyone else gets last year's model. And that has profound implications for education, work, and the economy.

The Frontier Isn't Coming Down

In previous technology cycles, early adopters paid a premium, then costs dropped as manufacturing scaled and competition increased. The first flat-screen TV cost $15,000. Within a decade, you could get one for $300.

AI isn't following that pattern. The frontier keeps moving upward:

• 2022: GPT-4 launches at ~$0.03/1K tokens — expensive, but one tier
• 2024: Model tiers emerge: Haiku (cheap), Sonnet (mid), Opus (expensive)
• 2025: OpenAI releases o1-pro at $150/1M — 10x the previous top tier
• 2026: Anthropic adds Capybara/Mythos above Opus; Google adds Gemini 2.5 Ultra above Pro

The pattern is clear: the best AI is becoming more expensive, not less. Budget models get cheaper (GPT-5 Nano is $0.05/1M), but the frontier keeps climbing.

Why AI Pricing Works Differently

Three factors break the "technology gets cheaper" rule for frontier AI:

1. Computing costs scale with capability. Training GPT-4 required ~25,000 A100 GPUs running for months. Training GPT-5.4 or Claude Mythos requires far more compute. The models don't just get "better" — they require fundamentally more resources to run.

2. There's no manufacturing efficiency gain. Unlike chips or screens, AI inference can't be "manufactured" more cheaply at scale. Each query still requires GPU compute. You can optimize at the margins (prompt caching, batching), but the fundamental cost structure is fixed.

3. Competitive pressure pushes capability, not price. OpenAI, Anthropic, and Google are racing to build better models. The competitive frontier is capability, not cost. They'll cut prices on older models while raising prices on new ones.

"We're in the hard takeoff. Right now. I go to sleep, there's some massive AI breakthrough, and when I wake up, there's another one." — Elon Musk, March 2026

The Inequality Multiplier

This creates something new: an intelligence divide based on purchasing power.

The consequences ripple through society:

For education: Students at well-funded universities get access to frontier models through institutional licenses. Students at underfunded schools get budget-tier AI — if they get any at all. The quality of AI-assisted learning becomes stratified by wealth.

For small businesses: A startup can access GPT-5 Nano for basic automation. But Claude Mythos-level cybersecurity analysis, code review, and research synthesis is reserved for companies that can afford $10K+ monthly API bills.

For developing nations: A March 2026 ILO/World Bank whitepaper warned that developing countries face AI-driven job losses before they can capture productivity gains. The reason: workers in jobs vulnerable to automation are already online, while those who could benefit from AI tools often lack internet access.

"Workers in jobs vulnerable to automation are often already online, even in low-income settings, meaning job losses could happen relatively quickly. These jobs often represent relatively higher-quality jobs in lower-income countries... AI-driven automation could close off these pathways." — ILO/World Bank whitepaper, March 2026

Microsoft's $50 Billion Response

In February 2026, Microsoft pledged $50 billion to tackle AI inequality, warning of a "growing divide" between those with access to frontier AI and those without. The company's AI Impact Summit highlighted three priorities:

• Digital infrastructure for developing nations
• AI literacy programs for underserved populations
• Affordable access to AI tools for education and small business

But $50 billion over "several years" is a fraction of what frontier AI companies spend quarterly. Microsoft's own Azure AI services charge premium prices for frontier models. The company is both warning about inequality and benefiting from it.

What New Zealand Should Do

New Zealand sits in an unusual position: wealthy enough to afford frontier AI, small enough to act quickly, but far from the centers of AI development. Several policy responses make sense:

1. Education programs should teach with frontier AI

The UK's AI Skills Boost program trains 10 million workers with free courses from Google, Microsoft, and IBM. But these courses teach foundational AI literacy — how to use ChatGPT for drafting text, how to prompt effectively.

They don't teach with Claude Mythos or GPT-5.4. The most capable models are reserved for those who can pay.

A meaningful NZ education initiative would:

• Negotiate institutional licenses for frontier models at universities and polytechs
• Train students on the tools they'll actually use in high-value work, not just budget-tier versions
• Create AI apprenticeships where students work with frontier models on real problems

2. Government should subsidise frontier AI access

Just as NZ subsidised broadband rollout to close the digital divide, it could subsidise frontier AI access for:

• Small businesses (< 20 employees) developing AI-assisted products
• Students and researchers at public institutions
• Non-profits and community organisations

The goal isn't free AI for everyone — it's ensuring the capability gap doesn't become an insurmountable wealth gap.

3. NZ needs its own AI infrastructure

The Southland data centre (consuming 6% of national power) will train models for international companies. But NZ has no sovereign AI capability — no models trained on NZ data, for NZ problems, with NZ values.

Building domestic AI infrastructure would:

• Reduce dependence on US AI companies that raise prices annually
• Create NZ-specific models (te reo Māori, NZ law, local context)
• Provide a fallback if international access becomes restricted

The Hard Truth

Technology optimists argue that AI will eventually become cheap enough for everyone. They point to GPT-5 Nano at $0.05/1M tokens as evidence that good AI is getting cheaper.

But "good enough" is the trap. GPT-5 Nano can summarise text. Claude Mythos can find vulnerabilities in your codebase that a human auditor would miss. One is a convenience; the other is a competitive advantage.

The question for New Zealand, and for every country, is whether to accept this divide or to actively intervene. The UK is training 10 million workers. Microsoft is investing $50 billion. The companies building frontier AI know the stakes.

Do we?