TypeSafe AI is a US-based AI lab that came out of stealth on September 15, 2026 and opened early access. Its model, Jev, is designed in the opposite direction from conversational large language models. Rather than producing prose for people to read, it focuses on returning a “decision” that software can use directly. It handles the judgments that occur constantly inside an application, such as “which department should this inquiry go to,” “on a scale of 100, how urgent is this post,” or “is it safe to run this operation automatically.” What comes back is not free-form text but a value that fits a predefined type, along with a probability and a confidence level indicating how certain it is.
Key Features
- Jev, a decision model that does not generate text: An architecture named System One Model that has no text-generation stage at all. Output is limited to typed, probabilistic decisions
- Three question primitives: Choice (pick from options), Score (assign a number), and Noul (return a yes/no probability). With only three forms to learn, there is less to figure out when designing prompts
- Calibrated confidence: Every answer carries a confidence score. You can implement “handle automatically when confidence is high, send to human review when it is low” simply by setting a threshold
- Fast responses through parallel sampling: Instead of generating tokens one at a time, it produces the entire output in parallel in a single query. The company states response times in the range of 70 to 500 milliseconds
- REST API and official SDKs: In addition to the REST API at
api.typesafe.ai, official client SDKs are available for Python and JavaScript
Pricing
| Plan | Price | Details |
|---|---|---|
| Usage-based (Jev / Early Access) | $0.042 per million input tokens | Output tokens are not billed. Available via API and SDKs |
There are no fixed tiers such as Free or Pro ─ billing is usage-based on the input token rate alone. The site has no standalone pricing page, and the figure above comes from the launch blog post. Because this is early-access pricing, it may change when general availability begins.
Pricing information is current as of September 2026. Please check the official site for the latest pricing.
Pros & Cons
✅ Pros
- You can embed background judgments without hurting perceived responsiveness, making it easier to design flows that do not need a loading spinner
- Since output is not billed, cost estimates depend only on input tokens even when you run judgments at high volume. Predictable cost makes it easier to decide on production use
- Return values are guaranteed by type, so you do not need retry logic or cleanup code for failed JSON parsing. That means less code and fewer places for bugs to hide
- By changing the confidence threshold, you can tune how much is automated while the system is running. You can start with “a human reviews everything” and widen automation as results come in
⚠️ Cons
- It cannot be used for text generation, summarization, or open-ended conversation. Its scope is limited to judgments expressible as Choice, Score, or Noul
- It can return a judgment that is valid as a type yet semantically wrong. Outside reviews have noted this, so handling confidence and preparing evaluation data remain the user’s responsibility
- The performance multiples cited for speed and cost advantages come mainly from the provider’s own benchmarks; independent third-party verification has not been confirmed
- The service only came out of stealth in September 2026 and is currently in waitlist-based early access. Terms such as a commercial SLA are not yet settled
Comparison with Similar Services
| Service | Difference |
|---|---|
| OpenAI Structured Outputs / JSON mode | A feature that retrofits existing large language models to enforce output conforming to a JSON schema. Jev differs in being a dedicated architecture that does not generate text at all |
| Instructor (Python library) | A library that wraps large language model calls and parses the result into a Pydantic schema. TypeSafe AI replaces the foundation model itself |
| Guardrails AI | A validation layer that checks and corrects large language model output. Jev handles the generation of decisions rather than their validation |
Who Is It For
- Developers who need to run judgment tasks at high volume, such as routing inquiries or scoring content
- People who have integrated a large language model but are spending time on slow responses and failed JSON parsing
- Business system owners who want to replace rule-based logic for fuzzy conditions that if statements cannot express
- Teams that want to tune the boundary between automated handling and human review numerically while in operation
- Anyone troubled by accumulating output-token costs who wants to split off at least the judgment workloads
Summary
Even as conversational AI has become commonplace, there are plenty of situations where software needs a judgment rather than prose. TypeSafe AI narrows in on exactly that, returning answers in a form machines can work with: types and confidence levels. If you are running a large volume of judgment tasks, the approach is worth trying. That said, it is at an early-access stage right after launch, and the performance evidence rests mainly on the provider’s own benchmarks. A realistic path is to evaluate it on a small scale with your own data, confirm the confidence thresholds, and only then put it into production.