TypeSafe AI, founded by ChatGPT and RLHF co-creator Diogo Almeida, released Jev, a transformer that outputs calibrated probabilities instead of text. Its free output tokens and billion‑metered inputs target high‑volume classification, potentially pressuring LLM API economics while demanding careful calibration testing. Developers report major classifier speedups, but independent validation remains early.
What Happened
On September 18, 2026, TypeSafe AI announced Jev, a transformer‑based model that does not generate text but instead produces calibrated probability scores for user‑defined labels. The release follows Diogo Almeida’s departure from OpenAI two years ago, where he helped build ChatGPT and co‑invent reinforcement learning from human feedback.
Almeida told TechCrunch that “We have lightning in a bottle, and yet it is not useful,” and added, “I’ve been battling that problem since then. It took me a while to come to the conclusion: The problem is we are optimizing for human language … We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language.”
Jev’s architecture is designed to be cheap and fast: output tokens are free, and input tokens are metered by the billion instead of the million. Because users define the output space in advance, the model cannot hallucinate prose, though it can still misclassify with high confidence.
Demand for Jev was immediate, and TypeSafe briefly lost the ability to serve users from its API while the developer community rushed to try it out.
Vercel, a company building agentic infrastructure, had used OpenAI’s ChatGPT Luna 5.6 to run a safety classifier. After swapping Luna for Jev, Vercel reported a significant speedup and improved accuracy, according to software engineer Pranit Sharma.
Sharma shared the results publicly:
We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev.
tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice pic.twitter.com/3G8tpRz7AG
— Pranit (@fazxes) September 16, 2026
Bryo AI CTO Nikhil Mudholkar ran his own test, pitting Jev against Gemini on a classification task.
Jev lost to Gemini on our email classification benchmark.
I’m still interested in putting it into production.
We tested 1,565 German and English business emails across 10 categories from Industrial suppliers.
The interesting result wasn’t accuracy. It was where the mistakes… https://t.co/lGhhM95H8U pic.twitter.com/29ARYcRrJg
— nikhil mudholkar (@nikhilmudholkar) September 17, 2026
What This Means For You
If your team uses an LLM as a classifier, router, moderation filter, or triage step, that is precisely the workload Jev targets. The pitch is narrow: fixed labels in, probabilities out. That narrowness quietly covers a large share of production AI calls.
Start with a shadow deployment. Run Jev beside your current model on real traffic and compare accuracy and p95 latency before you change anything. Vercel’s speedup is a useful reference point, but your data distribution may behave differently.
Study the pricing shape, not just the headline. Free output tokens and billion‑metered inputs reshape the math for high‑volume, low‑variance tasks. Long‑context generation, summarization, and chat remain LLM territory, so do not cancel those budgets yet.
“Cannot hallucinate” deserves scrutiny. A fixed label set removes invented text, not wrong answers. Calibration drift, distribution shift, and weak training data can still produce confident errors. Ask for calibration metrics computed on your data, not just accuracy scores from a vendor’s benchmark.
Auditability may genuinely improve. Probabilistic outputs over defined labels are easier to log, threshold, and explain to a regulator than free‑form prose. That matters in regulated pipelines such as insurance underwriting, where you must justify a decision to an examiner.
Apply the same logic to clinical triage or patient‑routing tools, where speed and traceability often beat conversational fluency. Teams building AI in telemedicine and telehealth should treat decision models as a separate evaluation track from their chat assistants.
Watch vendor risk carefully. TypeSafe’s capacity crunch is a warning about depending on a young startup. Negotiate rate limits, uptime commitments, and fallback routing before you place Jev on a critical path. Keeping an LLM fallback live is cheap insurance against a single point of failure.
Finally, revisit your evaluation harness. If your evals only measure text quality, they will not tell you whether a decision model is well calibrated. You need labeled test sets, threshold sweeps, and drift monitoring in place first.
Why It Matters
This suggests the LLM era is narrowing rather than ending. Language models won the interface war, but a large slice of enterprise automation never needed language at all.
If Almeida is right, the industry spent four years optimizing a capability that computers never required for most backend work. Specialized decision models could absorb the high‑volume, low‑margin inference that currently consumes the majority of AI spend.
That could pressure incumbent API economics. Cheaper classifiers mean less revenue per automated task, and pricing pressure has a habit of spreading to adjacent workloads.
It also connects to a broader reckoning with reliability and liability exposure across the AI industry. Auditable, bounded outputs are easier to defend than a chatbot that improvises, which may matter more to buyers than raw capability.
Treat the enthusiasm as promising but early. The performance claims here come from developers testing a model released this week, not from independent, longitudinal benchmarks.
Key Takeaway
- TypeSafe AI, founded by ChatGPT and RLHF co‑creator Diogo Almeida, released Jev — a transformer that outputs probabilities instead of text.
- The economics are inverted: output tokens are free, input tokens are billed by the billion, and no text means no hallucinated prose.
- Vercel reported replacing ChatGPT Luna 5.6 with Jev for a safety classifier and getting significantly faster results with better accuracy.
- Pilot Jev on narrow classification tasks under shadow traffic first, and keep an LLM fallback given TypeSafe’s API capacity strain.
Frequently Asked Questions
Is Jev a large language model?
No. It is transformer‑based but produces probabilities the company calls “calibrated decisions,” with outputs defined in advance by the user.
Does “cannot hallucinate” mean it is always correct?
No. It cannot invent text, but it can still assign a wrong label with high confidence. Calibration drift and distribution shift remain real risks you should monitor.
Should you replace your LLM classifier right away?
Not blindly. Run a shadow test on your own traffic, compare accuracy and latency, and only migrate once your evaluation harness and fallback route are ready.
Leave a Reply