TypeSafe AI’s Jev Model Replaces Language With Probability Outputs

When Almeida left OpenAI two years ago, he carried with him expertise in developing the techniques that powered modern large language models. Yet despite witnessing their remarkable capabilities, he identified what he saw as a fundamental misalignment: the entire industry had optimized for human language output, not for the actual work automation required.

“We have lightning in a bottle, and yet it is not useful,” he told TechCrunch. That disconnect—between what current AI does well and what software actually needs—became the founding premise for TypeSafe AI.

This week, TypeSafe unveiled Jev, a transformer-based model that operates on an entirely different principle. Rather than generating language, it outputs probabilities. The company calls these “calibrated decisions.” This seemingly small shift in approach creates profound practical differences.

A Model Built for Automation, Not Language

Almeida’s observation at OpenAI proved prescient: most recent progress centered on language because that was optimized for human use. But machines don’t speak human language, and forcing them through that interface introduces both inefficiency and risk.

By rejecting text generation entirely, Jev gains several properties that large language models inherently struggle with. It runs fast and costs dramatically less to operate than alternative solutions. Since the outputs are predefined—developers specify what possible answers Jev should choose between—the model cannot hallucinate. There’s no risk of it returning nonsensical data.

Pricing and Performance at Scale

The cost structure differs fundamentally from how LLM pricing works. While most language models charge by input and output tokens, Jev charges only for input tokens, metered at the billion-token level rather than the million. Output tokens are free. This creates a dramatically different economics equation for developers deploying the model repeatedly across large workloads.

The cost advantage proved so compelling that demand briefly exceeded TypeSafe’s infrastructure capacity. The company temporarily lost the ability to serve users through its API, a sign that developers saw immediate practical value in what they’d been testing.

The Training Approach: Synthetic Data and Calibration

Almeida disclosed that Jev relies exclusively on synthetic data, created through what he calls “reinforcement learning from calibrated decisions.” The approach differs sharply from how most modern AI is trained. Rather than fine-tuning on human feedback, the model learns from data the company generates itself.

“We made an early bet that we will be making all of our data, and that has been one of the best bets I’ve ever made in my life—better than our launch, in my opinion, better than RLHF,” Almeida said. He described the synthetic data work as the company’s core research effort, representing half the organization.

Real-World Performance From Early Adopters

Read also: FAA deploys $875M AI system to address controller shortage · AI Labs Deploy AI Monitors to Track Rogue Agents · UN and Google Launch Data Platform Built for AI

The appeal became concrete when developers began testing Jev against established alternatives. Two significant case studies emerged from early usage.

Vercel’s Safety Classification Workflow

Vercel, which builds infrastructure for agentic AI systems, had previously used OpenAI’s ChatGPT Luna 5.6 to classify and review commands for safety. When the team switched to Jev, performance shifted dramatically. Pranit Sharma, a software engineer at Vercel, reported that the model delivered results 5 to 18 times faster than the LLM it replaced, while simultaneously improving accuracy.

That speedup matters in a production context. Safety checks that can run nearly an order of magnitude quicker enable higher throughput systems without sacrificing the security guarantees they’re meant to enforce.

Bryo AI’s Email Classification and Cost Analysis

Nikhil Mudholkar, CTO of Bryo AI, conducted a different benchmark: classifying business emails using both Jev and Google’s Gemini. In his testing, Gemini achieved slightly higher accuracy, but the cost difference overwhelmed that advantage. Jev proved 10 to 20 times cheaper.

More notably, Mudholkar highlighted a capability neither LLM users typically emphasize: Jev returns genuine probability scores. “It is the only one that hands back a real probability which makes it ideal for automating workflows!!” he noted. That confidence measure gives downstream systems a way to handle ambiguity—deferring to human review when a classification falls below a confidence threshold, or acting autonomously when confidence runs high.

Beyond Replacement: Augmentation and Monitoring

While Jev can directly replace LLMs in certain tasks, Almeida sees additional applications that could define the model’s long-term value. Using one AI system to monitor another—deploying agents to watch agents—quickly becomes expensive. But because Jev costs so little to run, it becomes practical to use it as a lightweight safety check. Almeida envisions deploying Jev to trace LLM agent execution, detecting potential jailbreaks or misbehavior before they cause harm.

The approach trades some responsibility to users. Armin Ronacher, CTO of Earendel (which maintains the open-source model framework Pi), explained the tradeoff: users must understand that low-probability outputs represent genuine uncertainty, not reliable answers. When Jev returns 95% confidence, users can act. At 50%, the outcome remains essentially a coin flip.

Another application Ronacher identified: model routing. Determining whether a particular task requires a specialized model or can run on a cheaper general system would normally require an LLM itself—expensive and slow. Jev’s speed and cost make real-time routing decisions feasible, directing work to the most appropriate model without expensive preprocessing overhead.

A Reference to Economics, and a Vision of Distributed Intelligence

The model’s name carries intentional meaning. William Stanley Jevons was a 19th-century economist famous for his paradox: as the cost of a commodity falls, usage increases rather than stays flat. Cheaper steel led to more steel being used everywhere, not less. Almeida drew the parallel deliberately. As the cost of deployed intelligence approaches zero, it should proliferate through software infrastructure, becoming ubiquitous rather than concentrated.

“We think that there’s just going to be smart software all over the place in a way that’s emergent and distributed … much more like the early internet than you know like the mega apps that people are trying to build right now,” Almeida said. His framing positions TypeSafe not as a competitor building the next flagship model, but as enabling infrastructure for a different computational paradigm.

The Competitive Landscape

For now, TypeSafe stands alone in offering this specific combination of properties. Armin Ronacher expects that to change quickly. “We should have seen this earlier in many ways, but presumably because the LLMs are so cheap and subsidized, you often don’t have to be creative yet,” he observed.

Once competitors recognize the market opportunity, the space will likely fragment into specialized models serving different task classes. TypeSafe itself is already planning new versions across different modalities, expanding the approach beyond text-based classification.

When asked whether TypeSafe qualifies as a “frontier lab,” Almeida pushed back against the framing. “The main product of Frontier Labs is fear or hype. I would like our main product to be intelligence,” he said, distinguishing his company from the high-profile research organizations that dominate AI discourse. The focus remains narrow: building practical tools for software automation, not pursuing open-ended research toward artificial general intelligence.