TypeSafe Jev
by TypeSafe AI
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification
TypeSafe Jev is a decision model that replaces LLM calls for classification, routing, scoring and yes/no judgments inside software and agent pipelines. Instead of generating text, it returns typed answers with calibrated probabilities in roughly 70-500 ms at $0.042 per million input tokens, aimed at engineering teams paying generative-model prices for simple decisions.
Jev is the first model from TypeSafe AI, a San Francisco startup founded in 2024 by CEO Diogo Almeida, a former OpenAI researcher credited as a co-inventor of RLHF, with co-founders Erik Gafni and Sasha Sheng. TypeSafe emerged from stealth on September 15, 2026 with a roughly $40 million seed round led by DCVC and launched Jev the same day as what it calls a 'System One model', after Daniel Kahneman's fast, intuitive mode of thinking. Jev is not a chat model and cannot write text or code. A developer sends a state (text or structured data, within a 64k-token request budget of which 32k covers the state plus the longest question) and a set of typed questions: Choice picks one of up to 255 options, Score places the state on a rubric scale, and Noul returns the probability that a statement is true. Jev scores every possible answer in a single non-autoregressive pass, trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions, so each answer is schema-valid and carries a probability; application code can act automatically at high confidence and escalate to a human or a stronger LLM when uncertain. Pricing is $0.042 per million input tokens with output free, end-to-end latency is 70-500 ms, and launch rate limits are 250,000 tokens per second and 1,200 requests per minute. Python and JavaScript SDKs and a REST endpoint are available, and Vercel exposed Jev through its AI SDK and AI Gateway, where VentureBeat reported about 13% of paid Gateway teams using it within 24 hours; LangChain, Langfuse and Cloudflare added integrations within three days. TypeSafe cleared 140,000 waitlisted developers in 36 hours and dropped the waitlist on September 20, 2026. The current version, Jev 1.13, is text-only and most accurate in English.
Platform or AI engineering leads running high-volume classification, routing or guardrail calls through generative LLMs, who are paying frontier-model prices for what are really yes/no or pick-one decisions.
Sub-second, schema-guaranteed decisions at a fraction of a cent each, with a calibrated confidence score that tells code when to act automatically and when to escalate.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based
- Target Market
- CTOs, Heads of AI, AI Platform Engineers, Enterprise Developers
- Deployment
- API-based, Cloud-only
- Founded
- 2024
- Headquarters
- San Francisco, United States
Key Features
- ✓Typed question primitives
Choice, Score and Noul question types always return schema-valid answers, eliminating JSON parsing failures and output-validation code in pipelines.
- ✓Calibrated confidence on every answer
Each answer carries a probability, so code can auto-act at high confidence and route uncertain cases to humans or stronger models.
- ✓Single-pass non-autoregressive inference
Scores all possible answers in parallel rather than generating tokens, delivering 70-500 ms latency versus seconds for generative LLM calls.
- ✓Input-only pricing
Charges $0.042 per million input tokens with output free, putting typical per-decision costs at a fraction of a cent.
- ✓Many questions over one shared state
Ask many independent questions about the same state in one request, enabling speculative fan-out patterns that cut both cost and latency.
- ✓SDKs and gateway access
Python and JavaScript SDKs, a REST API, and Vercel AI SDK and AI Gateway support with per-request zero-data-retention options.
Capabilities
Use Cases
- •Support ticket triage
Route incoming tickets by department, severity and refund intent in a single call, sending low-confidence tickets to human agents.
- •Agent tool-call approval
Classify an agent's proposed shell commands or tool calls as destructive and require human confirmation before anything is executed.
- •Document classification at scale
Label large document sets cheaply; one independent tester classified government hearing submissions for about $0.22 per 1,000 documents at 0.32 s median latency.
- •LLM output verification
Score whether a generated response meets compliance or quality criteria before it is shown to a customer, blocking failures early.
- •Model and intent routing
Score each request's intent and complexity upstream to decide whether a cheap model, a stronger LLM or a human should handle it.
Ideal For
Best For
- ✓High-volume support-ticket triage and intent routing where a full LLM call per ticket is overkill
- ✓Safety gates in agent pipelines, such as classifying shell commands or tool calls as destructive before an agent executes them
- ✓Labeling and classification pipelines that need a calibrated confidence to decide which items go to human review
- ✓Cheap first-pass filtering that sends only uncertain cases on to an expensive frontier model
- ✓Real-time interfaces that need sub-second judgments such as content filtering or tone detection
Not Ideal For
- ✗Anything that needs generated text, summaries or code: Jev only returns choices, scores and probabilities, so it must be paired with an LLM for writing
- ✗Tasks involving arithmetic, reliable counting, date comparison or precise measurement, which independent testers document as failure modes
- ✗Regulated buyers who need SOC 2 or HIPAA attestations, SSO or a self-hosted deployment today; none are published and the service runs from the US West Coast
- ✗High-stakes decisions on adversarial, user-controlled input without an extra guard, since injected text measurably shifts Jev's probabilities
Integrations
Deployment
Market Analysis
Pros
- ✓Very low cost and latency for decision tasks: $0.042 per million input tokens and 70-500 ms, corroborated by independent tests (0.32 s median, half a cent for 24 documents)
- ✓Calibration held up in independent testing: at 0.9+ confidence one tester saw 14 of 15 answers agree with reference labels
- ✓Schema-guaranteed outputs remove parsing and retry logic from agent pipelines
- ✓Unusually fast ecosystem uptake: Vercel AI SDK/Gateway, LangChain, Langfuse and Cloudflare integrations within days of launch
Cons
- ✗Cannot generate text or code, and independent write-ups list arithmetic, counting, date comparison, double negatives and vague scoring criteria as failure modes
- ✗Prompt injection influences verdicts: in one test adversarial input cut a command-block probability from 0.76 to 0.48, and Pydantic warns that reordering options can move the answer
- ✗Accuracy degrades with complex, multi-qualifier instructions and large irrelevant context; text-only, English-first, 64k-token cap
- ✗One week old with no named production customers, no disclosed revenue and no published security certifications; Forkast notes its benchmarks measure agreement with other models rather than ground truth
Pricing
Pay-as-you-go API
$0.042 per 1M input tokens (output free)
- ✓Jev 1.13 via REST, Python and JavaScript SDKs
- ✓64k-token context per request
- ✓250,000 tokens/sec and 1,200 requests/min rate limits (adjusting dynamically)
- ✓No training on customer data
Via Vercel AI Gateway
$0.042 per 1M input tokens (output free)
- ✓experimental_evaluate() in the Vercel AI SDK
- ✓Zero data retention and no-training options per request
- ✓Gateway logs, reporting and budgets
Enterprise zero data retention
Contact for pricing
- ✓Zero data retention on request via privacy@typesafe.ai
- ✓Data Processing Agreement and Master Customer Agreement
Jev is metered only on input: $0.042 per million input tokens ($42 per billion), with output tokens free, and the same price applies through Vercel AI Gateway. No free tier or published enterprise plan exists; zero data retention is available to enterprise customers on request, and TypeSafe warns that rate limits are adjusting dynamically under launch demand. One independent test put the cost at about $0.22 per 1,000 documents versus $1.31 for DeepSeek without reasoning.
Security & Compliance
Connect
Sources
This page was written from 12 sources, 10 on domains other than typesafe.ai.
- 1.typesafe.ai — typesafe.aivendor
- 2.typesafe.ai — introducing system one models and jevvendor
- 3.docs.typesafe.ai — models.md
- 4.docs.typesafe.ai — legal.md
- 5.venturebeat.com — companies are putting jev in charge of ai agent decisions an
- 6.flaviocopes.com — jev
- 7.lindfors.no — a first look at typesafes jev
- 8.forkast.news — typesafe ais jev is not an llm and that may be the point
- 9.theaiinsider.tech — typesafe ai emerges from stealth with 40m to build machine n
- 10.techstartups.com — typesafe ai an ai startup founded by chatgpt co inventor eme
- 11.vercel.com — typesafe jev and ai sdk
- 12.hn.algolia.com — search
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Xiaomi MiMo-V2.6
MIT-licensed, 1M-context multimodal MoE models that took the top open-weights score on Artificial Analysis at well under $1 per million tokens
PrismML Bonsai
Open-weight 1-bit and ternary LLMs that run 27B-class AI on laptops and phones
Sakana AI Fugu Max
Multi-agent orchestration behind one API — frontier-grade results at $2 per million input tokens
Arcee Trinity
US-built open-weight model family, from on-device Trinity Nano to the 400B-parameter Trinity Large, that you can run on your own infrastructure