One of the unexpected releases of Sept 2026 was Jev, a promptable classifier that took X and the enterprise world by storm. Within a few weeks, open source models were springing up and OpenAI, Cloudflare and Liquid had released their own models and endpoints as commercial offerings. OpenAI also made it one of the key announcements during their developer day.
Jev had also spread quickly through the enterprise with the TypeSafe team sharing that 25% of Fortune 500 is using Jev just a few weeks after launch. Cost, ease of use and hallucination free outputs are truly enticing, yet the question remains, why now?
Cheap classifiers have been around for a long time (in AI years). Classification models powered applications on the early internet and were further democratized on Hugging Face with the release of BERT and CNN based models. These model families, Hugging Face’s Transformers library, and the open sourcing of TensorFlow and PyTorch immensely simplified how a small team could deploy very capable AI models. A team, or even an individual with the right data and intermediate knowledge on fine tuning models could get good to great results quite quickly. However, there were still some challenges. ML and programming knowledge was still needed, infrastructure for training and inference must be available, and with every task or data evolution, models needed to be trained, tested and redeployed. In 2018-2022 we had a boom of MLOps and platform tooling to support this need, but it did not lead to the AI rocketship we saw after Nov 30, 2022; ChatGPT launch day.
Reflecting back on the past four years of ChatGPT, there are a few things that stood out about the initial ChatGPT launch and subsequent adoption:
- Simplicity of using the model, there was no ML or Software barrier that people often feel about interacting with AI, it was conversational and stood out compared to earlier GPT-3 versions
- A virality of easy-to-screenshot results that became the conversation topic of the Winter Holidays
- A free version to reduce the barrier of entry
As the holidays subsided, 2023 began as the year of demos, use case pondering and rushed releases. Demos went to production and software standards disappeared. Testing was done to make the demo look good and deployed to make sure a company could point to AI in their implementation. Eventually towards the end of 2023 and start of 2024, ROI, production stability and measuring model accuracy returned as core discussion topics as teams began to operationalize how AI was used. The core reason why AI adoption continued to grow exponentially remained the same; whether you wanted AI for Customer Support, Coding, Workflow Improvement or drawing pelicans, a simple API and UX was the differentiating factor.
This simplified experience can be broadly referred to as “AI as a Service”, and every month the term and experience continues to evolve. Forward Deployed Engineers and Consultants may deliver it for some organizations, and for others it might be the existence of a highly available and simple AI. For many in the enterprise space, the “as a Service” might be defined as getting governance and security buy-in, while for others it's allowing builders to focus on the customer rather than the underlying infrastructure.
In an AI boom where model capabilities change rapidly, the zeitgeist can often dictate roadmaps rather than the best way to solve the problem. Agentic tasks emerged with the capabilities of GPT-4 and continued to expand with the latest generations of the largest and most capable models. And yet as capabilities expanded, the frontier was not always necessary, with many smaller and simpler problems capable of accomplishing core business and personal tasks at a fraction of the cost. Model costs had been rapidly decreasing since GPT-3 until 2025 when Google began hiking prices for Gemini 2.0 generation, and OpenAI followed suit when this became consistent in 2026 with higher prices for 5.4 nano. This trend continued and we saw stories of Uber, Coinbase, Meta, Microsoft and others rapidly decelerating their token spend due to cost, resulting in many organizations moving to internal or open source models. AI Harnesses also continued to evolve as many coding and agentic tools combined multiple models together using a stronger orchestrator to delegate to subagents for simpler tasks, or single agents consulting with larger models in an advisor capacity.
Smaller models became capable enough to do 90% of tasks but frontier labs focused on pushing the frontier as they continued to race (pace?) to AGI.
OpenAI re-opened the cost discussion with reducing costs on Luna 5.6 unexpectedly and again for Luna 6, bringing it back to the price point of $0.10 per million tokens for inputs. In many ways the Luna 5.6 price changes saved many wrapper companies who were struggling with low margins and enabled an explosion of usage as seen on OpenRouter.
The cost pendulum was already swinging back to affordability and Jev launched with a $0.042 a million input token price, a fast response and easy to understand API. Some clever marketing fuelled the fire, but cost, simplicity and generalizability defined the trajectory of its launch. Beyond my anecdotal experience of continuing to train classifiers with success amid the generative experience, one of the first Jev blog posts The Bitterest Lesson highlights this idea of selecting the right problem/problem and data as the foundation of effective AI.
As the AI market continues to grow (with only 2.2% of US Adults paying for AI subscriptions currently based on PNC data), these three factors will guide adoption, ranging from personal agents to enterprise adoption. Frontier intelligence will pave the way for the future, but a genius in a datacenter is not always the most approachable or affordable figure.
Top comments (0)