TypeSafe AI emerged from stealth in September 2026 with a $40M Series Seed led by DCVC at an estimated $200M post-money valuation. Within ten days of launch, media reported that investors were discussing financing the company at a valuation around $10B, but the talks were not confirmed as a completed round.

TypeSafe was founded in 2024 by CEO Diogo Almeida, COO Sasha Sheng, and CTO Erik Gafni. Almeida previously worked on instruction-following and reinforcement-learning research at OpenAI and Google Brain, while the broader team includes alumni of Meta FAIR, Stripe, Airbnb, Plaid, and Docker.

TypeSafe AI builds "System One Models," a class of machine-learning models designed to make bounded decisions inside software rather than generate prose for people. Its first public model, Jev, accepts unstructured context together with typed questions and returns three decision primitives with probabilities and confidence estimates: Noul for yes-or-no questions, Choice for selecting among defined options, and Score for ratings along a scale.

The output design lets developers set thresholds for automatic action and route uncertain cases to a human or a more expensive model. A workflow can use Jev to classify inputs, extract typed fields, rank options, route requests, score risk, evaluate model responses, or combine many small decisions in code. TypeSafe recommends decomposing complex tasks into independent narrow questions, keeping deterministic rules and calculations in code, and using model probabilities to decide when to act or escalate. The service is exposed through an API and Python and JavaScript SDKs.

TypeSafe developed a new model architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions. Jev was trained primarily on synthetic data and is optimized for decision tasks rather than open-ended generation. The model does not return free-form text, which narrows the surface area for conventional text hallucinations but does not eliminate incorrect decisions.

TypeSafe says it does not train or fine-tune models on customer prompts or other API input. Deployed model versions are fixed rather than silently changed, allowing developers to pin behavior as TypeSafe releases new versions.

TypeSafe benchmarks Jev at 70 to 500 milliseconds on selected tasks and claims speed and cost advantages of up to 193.6 times and 444.6 times, respectively, against multi-step language-model workflows. Those figures vary by workload and reflect the narrower job Jev is designed to perform.

TypeSafe sells Jev through a usage-priced developer API. Public pricing is $42 per billion input tokens, or $0.042 per million, with no charge for output tokens. Because Jev emits compact typed decisions rather than generated passages, usage is concentrated in the input context and the number of decisions a customer makes.

Within a week of launch, Jev had surpassed 1 trillion input tokens per day, with traffic continuing overnight as automated systems called the API. The launch post generated 40 million impressions and TypeSafe received more than 150,000 waitlist signups in its first 24 hours.

Within a day of launching on Vercel's AI Gateway, Jev became the gateway's fastest-adopted model launch, reaching nearly 13% of paid teams—more than twice the adoption of any previous model launch.

The initial go-to-market targets software teams that currently use frontier language models, hand-built rules, or conventional classifiers for routing, moderation, extraction, scoring, and evaluation. Developers can start through self-service API access, while larger deployments can expand into contracted volume, security review, service-level commitments, private data handling, and support.

TypeSafe's economics depend on keeping inference cost below its low public price while generating enough decision volume to support infrastructure and research spending. The company can improve margins through model and sampler optimization, batching, caching, and hardware utilization. A successful deployment can be sticky because decision schemas, confidence thresholds, escalation logic, and monitoring become embedded in the customer's application.

TypeSafe competes with frontier language-model APIs, smaller task-specific models, classical machine learning, rules engines, and emerging decision-model products.

OpenAI, Anthropic, Google, Meta, and other model providers support structured outputs, tool use, classification, and evaluation alongside general text generation. Customers may prefer one broad model platform rather than add a specialized endpoint. OpenAI's Decisions API validates demand for machine-oriented model outputs while creating direct competition from a provider with larger distribution, compute resources, and an installed developer base.

AWS's open-source Strands Decider model gives developers a small model they can inspect, fine-tune, and deploy inside their own infrastructure. Fine-tuned open models and conventional classifiers can also be inexpensive at high volume when a task is stable and labeled data is available. TypeSafe must show that Jev's calibration, generality, latency, and ease of use justify an external paid service.

Rules engines remain predictable and auditable for well-defined decisions, while traditional machine-learning platforms can train models around proprietary customer data. TypeSafe is best positioned where inputs are messy, decision types change frequently, or teams lack enough labeled data to maintain a separate model for every task.

TypeSafe can expand from individual classification and routing calls into a control layer for software agents and automated business processes.

Autonomous agents need fast checks before they send messages, execute code, move money, disclose data, or call external systems. Jev can evaluate intent, policy compliance, confidence, and expected outcome at each branch, escalating ambiguous cases without routing every step through a larger generative model.

Fraud screening, trust and safety, support routing, document triage, lead scoring, content labeling, and transaction review each involve many bounded decisions over unstructured data. These workloads can generate recurring API volume and reward low latency and calibrated uncertainty more than fluent text generation.

TypeSafe can add monitoring, audit logs, model versioning, private deployment, customer-specific calibration, and workflow analytics around the core API. Those capabilities would move the product from a model endpoint toward an enterprise decision platform. Additional modalities could extend the same typed-decision interface from text into images, audio, video, and sensor data.

Early product and adoption risk: Jev entered early access in September 2026. Strong benchmark results and launch-week usage may not translate into sustained adoption across diverse real-world distributions.

Incumbent and open-source competition: Frontier model providers can add cheaper decision endpoints to existing platforms, while open models let customers self-host. Rapid launches from OpenAI and AWS immediately after Jev's debut show that the category can attract well-resourced competitors and may commoditize quickly.

Scope and accuracy limits: Jev intentionally does not generate text, so it cannot replace language models for many workflows. Typed output prevents malformed responses but does not guarantee that a decision is correct, and calibration can degrade when customer data differs from the model's training distribution.

Synthetic-data dependence: Synthetic training can scale coverage without collecting sensitive customer data, but it may reproduce assumptions from the models or rules that generated it and miss rare real-world cases. TypeSafe must continuously validate performance on customer workloads without compromising privacy.

Pricing and infrastructure risk: The public price is low enough that small changes in model size, accelerator utilization, or traffic mix can affect gross margin. If customers use Jev mainly as an inexpensive pre-filter and reserve valuable decisions for larger models, TypeSafe may need very high volume to support its research and serving costs.