AI is moving in two directions at once. Some companies are building systems that reason through business tasks, while others are designing models that avoid free-form text altogether. At the same time, leaders of major U.S. AI labs are warning that future systems could improve on their own and slip beyond human control.
Diogo Almeida, a researcher who helped build the methods that taught AI to talk to people, is now working on a very different kind of model. His next system cannot generate text at all, and that limit is part of its design.
Jev trades conversation for speed and control
TypeSafe launched Jev as an AI system that chooses between options set in advance instead of producing open-ended text. The company says Jev runs at $42 for a billion input tokens, while its output is free. Responses arrive in 70 to 500 milliseconds, making Jev 40 to 200 times faster than today’s LLMs.
That design also changes the usual concern about unreliable AI answers. TypeSafe says, “Jev can’t hallucinate, because it only chooses between options set in advance.” Since Jev does not generate text and cannot move outside its prepared choices, it avoids the kind of open-ended response that can produce invented information.
Almeida described the system as “more like a database than a coworker.” That comparison captures the trade-off: Jev is built for speed, predictable choices, and low cost rather than conversation. Its pricing is also listed as 238 times below Claude Fable 5.1’s pricing.
The model’s narrow design may make it useful for tasks where the possible answers are already known. It also shows how AI development is moving beyond a single goal of making systems talk more naturally. In Jev’s case, removing text generation is the feature.
Salesforce builds a synthetic reasoning model
Salesforce introduced Koa, an in-house reasoning model for its sales and support agents, along with AIforce. Koa’s training set is fully synthetic, which means it simulates personas without using customer data.
On an internal CRM benchmark, Koa made three times fewer errors than top models. That result places the focus on a practical question for business AI: how well can a system handle sales and support work without exposing real customer information during training?
Koa and Jev take different approaches. Jev limits its possible outputs to control speed and accuracy, while Koa uses synthetic training data for reasoning tasks connected to customer relationship management. Both systems show companies building models around specific jobs instead of treating one general chatbot as the answer to every problem.
Tools and warnings shape the next stage
Developers are also getting new ways to compare and use AI models. An OpenRouter guide explains how to add models to Codex and Claude Code, with steps for setting up API keys, comparing models, and testing different options. The guide puts model choice into the hands of developers who want to examine several systems for the same task.
StackAI offers a platform to build, deploy, and scale AI transformations. Together with the OpenRouter guide, it reflects a growing focus on the work around models: connecting them to tools, testing their results, and moving them into real processes.
That progress comes with a warning. A $200 million startup aims to fix AI’s overconfidence problem, while leaders of major U.S. AI labs have warned about AI improving on its own and slipping beyond human control. The heads of leading U.S. AI labs came together over the weekend to slow AI development.
Sam Altman, CEO and co-founder of OpenAI, spoke at the 2026 Dreamforce conference in San Francisco on Tuesday, September 16, 2026. Dario Amodei, head of Anthropic, and Elon Musk, head of xAI, are also part of the group of leaders connected to the debate over how fast AI should advance.
The concern is tied to how quickly the technology has changed. AI has moved from the hallucination-prone ChatGPT of 2022 toward recursive self-improvement, where a system could help improve future versions of AI. That possibility explains why speed, reliability, and control now sit at the center of the same conversation.
Jev represents one answer: limit the system’s choices. Koa represents another: train a reasoning model on synthetic personas and measure its errors on a business benchmark. The warnings from OpenAI, Anthropic, and xAI point to the unresolved question behind both approaches—how much control people will keep as AI systems become more capable.
Based on
