Editorial illustration for Vercel Engineer Uses OpenAI's ChatGPT Luna 5.6 to Review Commands
ChatGPT Luna 5.6 Now Reviews Code Commands Automatically
Vercel Engineer Uses OpenAI's ChatGPT Luna 5.6 to Review Commands
Diogo Almeida spent four years building the technology behind ChatGPT before deciding it wasn't good enough. As one of the OpenAI researchers who helped invent reinforcement learning from human feedback, the training method that made modern chatbots possible, Almeida had a front-row seat to the industry he helped create. Two years ago he left to start TypeSafe AI, betting that the entire approach of making machines talk like humans was a dead end for one of AI's most obvious use cases: automation.
This week that bet produced Jev, a transformer-based model that skips language entirely. Instead of generating text, it spits out probabilities, what TypeSafe calls "calibrated decisions." Developers define the possible outputs ahead of time, which means the model can't hallucinate in the way chatbots do, and because it isn't wrestling with grammar and syntax, it runs far cheaper and faster. Output tokens are free. Input tokens get billed by the billion rather than the million.
The response from developers has been loud enough that TypeSafe's API briefly buckled under demand. What they're finding is a model built less for conversation than for the plumbing behind it.
Two years ago, Almeida left OpenAI to start TypeSafe AI, a startup trying to fix that problem. This week, the company released a new transformer-based model, Jev, that is not a large language model (LLM). It doesn’t output text, but instead produces probabilities, or what the company calls “calibrated decisions.”
Why this matters
Sharma's number, 5 to 18 times faster, is the kind of gap that should make anyone running safety classifiers at scale stop and check their own benchmarks. Command review is exactly the sort of narrow, high-volume task where latency compounds fast: every millisecond a classifier spends deciding whether to block an action gets multiplied across thousands of agentic calls a day. If Jev really holds that speed advantage without giving up accuracy, teams building agent infrastructure now have a concrete reason to test alternatives to OpenAI's models for these guardrail jobs, rather than defaulting to whichever provider trained the flashiest chatbot.
Almeida's framing is the more interesting thread here. His complaint isn't that ChatGPT is weak, it's that RLHF optimized for something other than raw usefulness, and he's been trying to fix that since leaving OpenAI. For founders and researchers, that's worth sitting with: the model that wins mindshare and the model that wins a specific production task, like Vercel's classifier, are turning out to be different products entirely.
Related Reading
Browse all OpenAI news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source
Related Topic OpenAI Developer: AI Agent "Swarms" Waste Tokens, Create "Coordination Tax
AI Tools & Apps
From Archives WGA East Executive Director Condemns Tech Oligarchs Pushing AI
From Archives Anthropic's Claude Hacked OpenAI Systems in 72 Hours Via Forum Flaw
From Archives OpenAI Exec Calls AI Training "Astonishing Theft" in Leaked Message
From Archives OpenAI Targets Legal Market With Astra's 230 Million Case Law Index
Policy & Regulation
Common Questions Answered
Why did Diogo Almeida leave OpenAI to start TypeSafe AI?
Almeida left OpenAI because he believed the approach of making machines talk like humans through large language models was a dead end for automation use cases. After spending four years building ChatGPT technology and helping invent reinforcement learning from human feedback, he decided the entire methodology wasn't suitable for practical automation tasks.
How is TypeSafe AI's Jev model different from traditional large language models like ChatGPT?
Jev is a transformer-based model that does not output text like traditional LLMs. Instead, it produces probabilities and what TypeSafe AI calls 'calibrated decisions,' making it fundamentally different in its approach to solving automation problems.
What performance advantage does Jev offer for command review and safety classifiers?
According to the article, Jev is 5 to 18 times faster than existing solutions for command review tasks. This speed advantage is particularly valuable for safety classifiers operating at scale, where latency compounds across thousands of agentic calls per day, and every millisecond of processing time has significant cumulative impact.
Why is the speed improvement of Jev significant for teams building agent infrastructure?
The speed improvement matters because command review is a narrow, high-volume task where latency multiplies quickly across numerous operations. If Jev maintains its speed advantage without sacrificing accuracy, it could substantially improve the efficiency and responsiveness of agent-based systems that rely on rapid decision-making.
Further Reading
- GPT 5.6 Luna API, Pricing & Playground | Vercel AI Gateway - Vercel AI Gateway
- GPT 5.6 Luna (Fast) API - AI Gateway - Vercel AI Gateway
- Finding GPT-4's mistakes with GPT-4 - OpenAI - OpenAI
- John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & ... - Dwarkesh Podcast
- Reinforcement Learning from Human Feedback - arXiv - arXiv
