AI Costs are Rising, But Not How You Think

Sam Altman once predicted that intelligence would become 'too cheap to measure,' a stark contrast to the current reality where AI costs are escalating.

In 2025, the creator's personal AI usage averaged $800 to $1,000 monthly, but in 2026, daily usage for tokens alone can exceed $1,000, driven by new models like Fable and Astra that generate more tokens at higher prices per token, leading to significantly increased operational costs for businesses.

The creator argues that AI costs have indeed decreased, but the conventional methods of measuring these costs are fundamentally flawed.

This analysis aims to clarify the discrepancy between the initial promise of cost reduction and the current expensive reality observed with advanced models like Astra.

Comparing the token pricing models of different generations, GPT-5 Medium cost $10 per million tokens out and $125 per million tokens in.

In contrast, GPT-6 Astra costs $50 per million tokens out and $10 per million tokens in, representing a 5x to 10x increase in both input and output token costs, a direct contradiction to expected cost reductions.

Evaluating AI expenses solely on token cost is insufficient, as the actual cost involves more complex factors.

A comprehensive assessment requires considering the cost per token, the number of tokens needed to complete a task, and the variety of tasks a model can perform, with 'intelligence per dollar' emerging as the ultimate metric for value.

While flagship AI models have seen an increase in token costs, the market has also introduced highly capable and significantly cheaper alternatives.

For example, models like GEM 5.3 Flash offer remarkable performance for agentic tasks at approximately one-thirtieth the cost of GPT-5, broadening the accessibility of powerful AI even as top-tier models become pricier.

Models such as Gemini 3.8 Flash, despite offering a lower per-token cost, can incur higher overall task expenses due to excessive token consumption.

This high token waste can significantly reduce the real-world cost advantage, narrowing the gap between cheaper and more expensive models from over 10x to under 3x.

The efficiency of a model is task-dependent; input-heavy operations prioritize low input token costs, while agentic tasks depend more on reliability and cache efficiency.

Achieving incremental improvements in AI intelligence often demands exponentially higher costs, where a slight score increase from 45-46 to the 50s can triple the financial outlay for a task.

The model is notably more cost-effective, being 21 times cheaper for input tokens and 55 times cheaper for output tokens compared to GPT-4, achieving approximately a 20x improvement in the cost-to-performance ratio, as tracked by Artificial Analysis for intelligence levels and task cost improvements.

From November 2022 to July 2024, the cost of sub-10 point intelligence tasks plummeted from 60 cents to 2 cents, demonstrating massive price reductions.

Further intensifying competition is evidenced by the o1-preview costs, which fell from $15 in September 2024 to just 16 cents through Grok 3 Mini by February 2025, indicating an accelerating trend of price drops as new, more affordable models quickly enter the market.

The cost for 20-point intelligence also experienced a dramatic decline, dropping from $155 to $26, and further to mere pennies with the introduction of even newer models.

The cost for new intelligence levels no longer experiences the exponential spikes seen previously.

Battlegrounds, particularly at the 40-point intelligence level, showcase newer models delivering higher scores at significantly reduced prices, with prices for a specific intelligence level dropping far more than 10x year-over-year.

AI model usage has evolved considerably, transitioning from basic autocomplete functions to comprehensive, end-to-end task execution.

This shift means models now handle larger, more intricate tasks with reduced human intervention, resulting in higher AI bills that reflect greater consumption and more tokens generated per prompt, rather than an increase in unit prices.

Models have become both cheaper and more capable, paradoxically leading to an overall increase in total spending on AI, a phenomenon compared to the historical impact of steam power.

Expenditure is rising not because the models themselves are more expensive, but because their enhanced capabilities and affordability encourage more extensive utilization across a wider range of applications.

A direct cost comparison, such as $0.25–$0.50 per million tokens for Flash versus $10 per million tokens for GPT-4, highlights Flash's efficiency and superior capability, underscoring how small models with advanced tool-use features now considerably outperform previously leading top-tier models.

The speaker acknowledges a significant error in a past prediction made two years prior, where he incorrectly suggested that AI improvements had reached a plateau.

This mistaken assessment stemmed from an over-reliance on inadequate benchmarks that failed to accurately capture the true extent of progress, leading the speaker to express disbelief at the rapid and cost-effective evolution of AI capabilities since his earlier prediction.

The promise of 'intelligence 10 times cheaper' has indeed materialized, making tasks performed last year significantly more affordable today.

The rising AI expenditure signifies a boost in productivity and the ability to tackle larger, more complex challenges, as current advanced workflows, even with high daily costs, would have been unfeasible or far less efficient with previous model generations.