The Classifier Renaissance and ChatGPT's Success Code: The Logic of Simplicity in AI Adoption

In September 2026, a promptable classifier called Jev unexpectedly went viral. Within just a few weeks, open-source models proliferated, and OpenAI, Cloudflare, and Liquid successively released their own models and commercial endpoints—OpenAI even made it a highlight of its DevDay. The TypeSafe team revealed that within just weeks of launch, 25% of Fortune 500 companies were already using Jev. Low cost, ease of use, and hallucination-free output—these advantages are enticing enough. But the real question worth asking is: why now?

Cheap AI Has Always Existed, But Adoption Never Exploded

Classification models are nothing new. As far back as the early internet era, classification models powered a massive number of applications; after Hugging Face released BERT and CNN-based models, this capability became further democratized. The open-sourcing of the Transformers library, TensorFlow, and PyTorch meant that a small team—or even an individual—with the right data and a moderate level of fine-tuning knowledge could quickly achieve decent or even excellent results.

But barriers remained. You needed machine learning and programming knowledge, infrastructure for training and inference, and every time you tackled a new task or the data evolved, the model had to be retrained, tested, and redeployed. Around 2021, there was a wave of MLOps and platform tooling to support this demand, but it didn't produce the rocket-like growth seen after ChatGPT's launch.

The answer lies in one word: "experience."

BERT (Bidirectional Encoder Representations from Transformers) is an open-source natural language processing model released by Google in 2018. Its core innovation lies in using a Transformer encoder to bidirectionally model the context of text, and it can be fine-tuned with a small amount of labeled data and then transferred to downstream tasks such as classification, question answering, and named entity recognition. BERT's emergence made "pre-training + fine-tuning" the standard paradigm for NLP, greatly lowering the barrier to building text classifiers. Subsequently, CNN-based text classification models were also widely used in sentiment analysis, spam detection, content moderation, and other scenarios. Hugging Face's Transformers library further unified model interfaces, allowing researchers and engineers to load pre-trained weights and complete fine-tuning in just a few lines of code. However, these tools solved the "training" problem without solving the "continuous operations" problem—data drift, task changes, and deployment operations still required professional team investment, which is the fundamental reason early classifiers struggled to achieve widespread adoption.

Three Key Differences ChatGPT Brought

Looking back at the four years since ChatGPT's release, several characteristics stand out:

Conversational interaction eliminated the ML and software barrier. Compared to early GPT-3, ChatGPT made users feel they weren't "operating AI"—they were just chatting.

Viral screenshots. Simple, screenshot-friendly results became a social topic during the winter holidays—a low-cost dissemination mechanism that no technical documentation could replace.

A free version lowered the barrier to entry. Free access turned curiosity into usage, and usage into habit.

After the holidays, 2023 became the year of demos, use-case deliberation, and rushed launches. Demos were pushed into production environments, and software engineering standards briefly disappeared. Testing existed only to make demos look good; deployment existed only so companies could point to AI in their reports. It wasn't until late 2023 to early 2024 that ROI, production stability, and model accuracy measurement returned to the core discussion.

But the core reason for growth never changed: whether you want AI to do customer service, coding, workflow improvement, or draw a pelican, a simple API and UX are the key differentiators.

Prompting differs from traditional supervised learning in a key way: classifiers no longer need to collect labeled data and retrain models for each task. Instead, classification targets are specified directly through natural language instructions or a few examples. This means task definition shifts from the "training phase" forward to the "inference phase." Data scientists no longer need to maintain multiple sets of model weights, and business users can describe requirements in natural language. "Promptable classifiers" like Jev are the product of this approach: users input category descriptions and text to be classified, and the model returns structured labels. Hallucination-free output benefits from the closed nature of classification tasks—the output space is constrained to predefined categories, and the model cannot generate free-form text, thereby avoiding the common fabrication problems of generative models.

"AI as a Service": Evolution from Demo to Enterprise-Grade Deployment

This simplified experience can be collectively called "AI as a Service," and it evolves every month. For some organizations, it may be delivered by forward-deployed engineers and consultants; for others, it means a highly available, simple AI endpoint. For many enterprises, "as a service" may be defined as obtaining governance and security approval; for other teams, it means letting builders focus on customers rather than underlying infrastructure.

In an AI boom where model capabilities change rapidly, the zeitgeist often dominates roadmaps rather than the best way to solve problems. Agentic tasks emerged with GPT-4's capabilities and continued to expand with the latest generation of large models. But as capabilities expand, the frontier is not always necessary—many smaller, simpler problems can accomplish core business and personal tasks at extremely low cost.

The Cost Pendulum: From Price Hikes to Renewed Affordability

Model costs fell rapidly starting with GPT-3, until Google began raising Gemini 2.0 generation prices in 2025, with OpenAI following suit. As this trend continued, stories emerged of companies like Uber, Coinbase, Meta, and Microsoft rapidly decelerating token spending due to costs, with many organizations turning to internal or open-source models (according to Forbes and The Information).

The AI Harness has also evolved: many coding and agent tools combine multiple models, using a stronger orchestrator to delegate simple tasks to sub-agents, or having a single agent consult a larger model as an advisor.

Small models are already sufficient for 90% of tasks, while frontier labs continue pushing boundaries for the AGI race. That was until OpenAI unexpectedly lowered Luna 5.6 pricing and reduced it again on Luna 6, bringing input prices back to the level of $0.10 per million tokens. Luna 5.6's price adjustment largely rescued many wrapper companies struggling with margins and triggered an explosion in usage on OpenRouter.

The cost pendulum has swung back to affordability, and Jev entered the market at $0.042 per million input tokens, with fast responses and an easy-to-understand API. Some clever marketing helped generate buzz, but cost, simplicity, and generality defined its trajectory.

Token is the basic unit of measurement for how large models process text. Typically, one English word corresponds to about 1 to 1.3 tokens, and one Chinese character corresponds to about 1 to 2 tokens. APIs price input and output tokens separately, so the price per million tokens directly determines the upper limit of large-scale deployment costs. For high-frequency call scenarios (such as customer service intent recognition and content moderation), input tokens often dominate, and drops in input pricing have a particularly significant impact on total costs. When input prices fall to $0.10 or even $0.042 per million tokens, the cost of a single classification call is already lower than the function call fees of many traditional cloud services, making it economically feasible to embed AI into every business process. Cost declines are not a linear trend, but a fluctuating process driven jointly by compute supply, model architecture optimization (such as sparsification and distillation), and market competition.

Choosing the Right Problem Matters More Than Choosing a Stronger Model

Beyond my personal experience of continuously succeeding at training classifiers, Jev's earliest blog post, "The Bitterest Lesson," pointed to the core: choosing the right problem and data is the foundation of effective AI.

According to PNC data, only 2.2% of American adults currently pay for AI service subscriptions. In this still-growing market, the three factors of cost, simplicity, and generality will continue to guide the path from personal agents to enterprise adoption. Frontier intelligence will pave the way for the future, but the geniuses in the data center are not always the most approachable and affordable characters.

The classifier renaissance is not a regression, but a recalibration of AI adoption logic: when model capabilities are cheap and easy enough, solving the right problem has more business value than chasing the largest model. ChatGPT's success code was never just intelligence itself, but packaging intelligence into a service anyone can use.

AI Harness (AI orchestration framework) refers to systems at the application layer that schedule, combine, and gracefully degrade multiple models or tools. Its typical patterns include: using a strong model as an orchestrator to distribute simple subtasks to cheap small models, or letting a main agent consult a larger expert model when needed. The value of a Harness lies in decoupling "model capability" from "task requirements"—not every step requires frontier intelligence, and routing and caching can significantly compress costs. In classifier scenarios like Jev, a Harness may manifest as a hybrid of rule engine and model: high-confidence results are adopted directly, while low-confidence samples are escalated to a stronger model for review. This layered architecture allows enterprises to balance accuracy and cost, and it is also the engineering foundation for delivering the "AI as a Service" experience.