TypeSafe AI has emerged from two years of stealth development with the public launch of Jev, a specialised model designed not to generate text, but to execute structured probabilistic decisions directly inside production codebases. The company was founded by Diogo Almeida, an OpenAI veteran and co-inventor of ChatGPT, and is positioning Jev as a purpose-built alternative to conversational language models for software systems that require fast, deterministic, automated logic.
What Is Jev and How Does It Differ from Standard Language Models
Unlike conventional large language models that generate sequential text tokens one after another, Jev takes unstructured state as input and returns type-safe structured values through a single parallel query. The model produces no text or string output at all. TypeSafe classifies it as a System One Model — a category of model focused on structured probabilistic decisions rather than conversational generation.
Almeida describes the model in direct terms: “Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.” By constraining outputs to predefined schemas, the system eliminates the parsing pipelines and guardrail wrappers that engineers typically bolt onto traditional language models when deploying them in business automation workflows.
The Training Method Behind Jev: RLCD Explained
TypeSafe built Jev around a proprietary training methodology called Reinforcement Learning for Calibrated Decisions (RLCD). This approach differs meaningfully from the methods used to train most frontier models. Conventional systems are trained using Reinforcement Learning with Human Feedback (RLHF), which optimises for human conversational preferences, or Reinforcement Learning with Verifiable Rewards (RLVR), which validates isolated programmatic checks.
RLCD is designed specifically to produce calibrated probabilities on execution logic, ensuring that the confidence scores Jev returns correlate directly with output accuracy. This makes the model’s uncertainty estimates actionable within production code, rather than decorative.
Jev’s parallel sampler is hardware-aware and evaluates all structured values simultaneously rather than generating them one token at a time. It handles high-cardinality selections of up to 255 discrete options using a two-stage process that scores options independently before final selection is made.
Speed and Cost Benchmarks Against Frontier Models
Performance measurements from TypeSafe’s West Coast servers show notable differences compared to conversational model deployments. Internal evaluations recorded end-to-end response latencies of 70 to 500 milliseconds for Jev, compared to 3 to 329 seconds for conversational frontier deployments.
Published workflow evaluations run across fixed compute graphs — measuring multi-step decision branching against consensus baselines from GPT-6 Astra and Fable 5.1 — recorded execution speeds up to 193.6 times faster.
Pricing reflects a similar gap. Input processing costs are set at $0.042 per million tokens, significantly below standard conversational model rates that range from $0.20 to $10 per million tokens. Because Jev generates structured states through parallel sampling rather than autoregressive token passes, TypeSafe does not meter output tokens, removing one of the primary cost variables associated with running language models at scale.
| Metric | Jev (TypeSafe) | Conversational Frontier Models |
|---|---|---|
| End-to-end latency | 70–500 milliseconds | 3–329 seconds |
| Input cost per million tokens | $0.042 | $0.20–$10.00 |
| Output token metering | None | Standard per-token billing |
| Speed vs. baseline models | Up to 193.6× faster | Baseline |
Real-World Testing: Doom Bots and Wikipedia Traversal
TypeSafe has released production demonstration results covering two distinct stress-testing scenarios. In the first, engineers operated a real-time reactive bot across Doom game state structures at 10 queries per second, with operating costs coming to approximately $7 per hour. The test validated Jev’s ability to resolve branching rules across high-speed, rapidly changing state inputs.
The second test used a Wikiracing format, evaluating link selection across dense encyclopaedia directory structures. Jev completed traversals in fewer steps than external non-reasoning models by selecting options without generating hallucinated dead-end links — a direct consequence of its schema-constrained output design.
Beyond these demonstrations, TypeSafe identifies the following as primary target deployment environments:
- Real-time feature extraction pipelines
- Petabyte-scale data workflows
- Output verification layers in automated systems
- Automated branching logic where hand-written rules become brittle over time
Access and Availability
TypeSafe opened early developer access on the day of the announcement and began onboarding engineering teams from its deployment waitlist. The company has not yet disclosed pricing tiers beyond the per-token input cost, or details about enterprise agreements. Given the model’s narrow but technically demanding use case — structured logic automation in production software — the initial focus appears squarely on engineering organisations rather than general-purpose users.
The launch positions TypeSafe in a distinct segment from general-purpose model providers. Where those companies compete on reasoning breadth and instruction following, TypeSafe is targeting the specific operational pain point of deploying reliable, fast, and cost-predictable automated decisions inside live software systems — a problem that has historically required significant engineering overhead to manage with standard language models.
Frequently Asked Questions
Jev does not generate text. Instead of producing sequential text tokens, it takes unstructured state as input and returns type-safe structured values through a single parallel query. This eliminates parsing pipelines and output hallucinations common in standard language model deployments.
TypeSafe prices Jev input processing at $0.042 per million tokens. Unlike conventional models, output tokens are not metered, since Jev generates structured states rather than autoregressive token sequences. Standard conversational frontier models typically cost between $0.20 and $10 per million input tokens.
TypeSafe AI was founded by Diogo Almeida, an OpenAI veteran who is credited as a co-inventor of ChatGPT. Almeida spent two years developing Jev in stealth before the company's public launch.
