Are these trend rates forecasts?

No. They are published estimates over defined historical windows and should not be treated as guaranteed future rates.

What does TokenLab model data add?

It shows current model availability, categories, and pricing units. External studies explain broader market movement.

Why do different sources disagree?

They measure different samples: public reports, platform traffic, benchmark tests, surveys, and model data. Keeping them separate makes comparisons clearer.

AI Model Reports

AI Model Industry Trends Report

Model scale, training cost, context length, enterprise spend, and current model-market composition.

Training compute growth

Epoch AI

4.5x/yr

growth rate

Compute used to train notable AI models has increased rapidly since 2010.

Frontier training cost growth

Epoch AI

3.5x/yr

growth rate

Epoch estimates frontier language-model training cost has climbed quickly since 2020.

Context window growth

Epoch AI

30x/yr

growth rate

Context windows have expanded extremely quickly since 2023.

Enterprise spend growth

Menlo Ventures

3.2x YoY

growth

Estimated enterprise GenAI spend grew from $11.5B in 2024 to $37B in 2025.

Industry trend indicators

Externally sourced signals for scale, context length, enterprise spending, usage-study size, and organizational adoption.

4.5x/yr

Training compute growth

3.5x/yr

Frontier training cost growth

30x/yr

Context window growth

$37B

Enterprise GenAI spend

3.2x YoY

Enterprise spend growth

100T tokens

OpenRouter usage study

88%

Organizational AI adoption

320x YoY

Reasoning-token growth

Scale is still the center of gravity, but inference economics are changing faster than raw model lists.

Epoch, Stanford, and Artificial Analysis all point to a market where compute, context windows, price, latency, and quality must be read together. Model names alone rarely explain fit.

Source: Trends in Artificial Intelligence / 2026 AI Index Report / Models: Intelligence, Performance & Price

Real usage is fragmented across consumer, enterprise, developer, and aggregator surfaces.

OpenRouter, OpenAI Signals, and Anthropic Economic Index each reveal a different slice of adoption. Read the sample and reporting window before comparing the numbers.

Source: State of AI: 100T Token Study / Signals Consumer Data / Anthropic Economic Index

Spending is moving from experiments into applications, but vendor rankings are not stable enough to freeze.

Menlo and a16z show adoption and spend shifting quickly. Dated source links matter more than evergreen winner claims.

Source: State of Generative AI in the Enterprise / Top 100 Gen AI Consumer Apps

Benchmarks are most useful when the chart explains what was measured.

Artificial Analysis and Arena expose ranking signals, but task mix, pricing assumptions, and sample design change how those rankings should be read.

Source: Models: Intelligence, Performance & Price / Arena Leaderboard