Pre-trained at scale on broad data. Adaptable across tasks without full retraining.
↑ All 10 Index models
Stage × Capability — plotting both axes
Where does every system actually sit?
X: He, Cao & Tan architectural stage. Y: DeepMind capability level (Morris et al., 2024). Hover any dot. Hit the button when you're ready.
Historical systems
Specialised / narrow AI
Innovare Index models
L1 Emerging — all current LLMs sit on this line
The finding — Stage 4 is not a tier
Same stage. 19× the price.
MiniMax-M3
$1.14
per million tokens · Cheap tier · Stage 4 · L1 Emerging
19×
price gap
Same stage. Same level.
Claude Fable 5
$22.00
per million tokens · Frontier tier · Stage 4 · L1 Emerging
He, Cao & Tan's stages describe architectural evolution across seventy years — how the field's approach changed each era. They don't rank competing products within the same era. Using "Stage 4" as a product tier is the most common misapplication of this taxonomy in vendor marketing.
Where the framework breaks
Two things the matrix won't tell you
The stage framework is genuinely useful for understanding what changed architecturally between eras. It breaks down in two places:
Hole 1 — Stage doesn't predict capability within an era
Fable 5 and MiniMax-M3 share a stage classification. They don't share benchmark scores, price, latency, or reasoning depth. If you're making a build-vs-buy decision, stage tells you almost nothing. You need the Index pricing data, the AA leaderboard, and task-specific evals.
Hole 2 — Hybrid products don't map cleanly
Palantir Foundry runs a Stage 2 statistical core with a Stage 4 AIP wrapper. The taxonomy assumes one architecture per system. Real enterprise AI products are layers — and the marketing tends to brand the wrapper, not the core. "Powered by AI" often means a Stage 4 face on a Stage 2 body.
Take this away
He, Cao & Tan’s four stages explain what changed architecturally between AI eras—but stage classification alone tells you almost nothing about which product to deploy or whether it will work for your specific task.
Name He, Cao & Tan’s four stages of AI development and what distinguishes each architecturally.
Stage 1 (Conceptual/Rule-based, 1950–1979): rules encoded by humans—ELIZA, early expert systems. Stage 2 (Statistical pattern matching, 1980–2011): systems find patterns in data humans pointed them at—Deep Blue, Watson. Stage 3 (Narrow deep learning, 2012–2016): neural networks learn representations autonomously—AlexNet, AlphaGo. Stage 4 (Foundation models, 2017–present): transformer architecture trained on internet-scale data, emergent multi-domain capability—GPT-4, Claude, Gemini.
Question 2 of 3
What does the Innovare Index data reveal when you plot all current frontier models on the Stage × Capability matrix?
Every model in the Innovare Index lands in Stage 4—but with a 19× price spread across the tier. The matrix confirms that stage classification cannot differentiate between products within an era. Two Stage 4 models share a classification but differ substantially in benchmark scores, price, latency, and reasoning depth. The matrix explains architectural history; it does not predict per-task capability.
Question 3 of 3
Why does the four-stage framework break down when you try to use it to make a deployment decision?
Two reasons. First, stage doesn’t predict capability within an era—all current frontier models are Stage 4, but they vary enormously on the tasks that matter for a specific deployment. You need Index pricing data, benchmark leaderboards, and task-specific evaluations. Second, hybrid products don’t map cleanly—many enterprise AI products layer a Stage 4 interface on a Stage 2 statistical core. The marketing brands the wrapper, not the core.
Sources
Four-stage framework
He, R., Cao, J., & Tan, T. (2025). Generative artificial intelligence: A systematic review. National Science Review.
DeepMind levels
Morris et al. (2024). Levels of AGI: Operationalizing Progress on the Path to AGI.arXiv:2311.02462.
Innovare Index
Pricing and AA rank data from Innovare Index, Edition 006, 1 July 2026.