Decision model API

How to use
TypeSafe AI.

Fast, cheap, calibrated yes/no, pick-one and rate-on-a-scale decisions inside software: routing, classification, scoring, guardrails and verification, where code keeps control flow and the model supplies the judgment.

Open TypeSafe AI ↗ Read the quickstart ↗
What it ismachine-native decision model API (System One Models) Workflows3 UpdatedSeptember 2026

TypeSafe AI is a machine-native decision model API (System One Models). Fast, cheap, calibrated yes/no, pick-one and rate-on-a-scale decisions inside software: routing, classification, scoring, guardrails and verification, where code keeps control flow and the model supplies the judgment. This guide covers the whole path in one place: official access, a first session that produces something reviewable, the checks that make output trustworthy, and the permissions worth limiting before you connect real work.

Troiana principle

AI should make the work easier to inspect. If the workflow removes the source, the owner, or the review step, redesign the workflow.

01

Access & setup

Find, install, and sign in to TypeSafe AI

Get into the official TypeSafe AI experience with the right account and a setup you understand.

  1. 01

    Start at https://typesafe.ai/ and confirm the domain before entering account or payment information.

  2. 02

    Availability: TypeSafe AI is accessed through the web console and Playground at console.typesafe.ai, the HTTP API at api.typesafe.ai, and official Python and JavaScript SDKs. Early access: developers are admitted from a waitlist.

  3. 03

    A TypeSafe account and API key, network access, and Python 3.10+ (typesafe-sdk), Node.js 20+ (@typesafe-ai/sdk), or any HTTP client. Text input only: strings, JSON objects, and arrays. No images, audio, or video yet.

  4. 04

    Sign in with the account you intend to keep using, then review plan, data, notification, and permission settings.

  5. 05

    Run one low-risk test task before connecting sensitive files, repositories, or workspace data.

  • Use official download pages.
  • Review permissions during setup.
  • Keep installers and applications updated.
02

First session

Your first useful TypeSafe AI session

Learn the interaction loop using a small task with a clear outcome.

  1. 01

    Open the Playground at console.typesafe.ai, paste one real piece of text as the state (a support ticket, a review, a log line) and add a single Noul question such as "Does this message express urgency?". Read the probability that comes back before adding anything else.

  2. 02

    Add a Choice question with three or four named options and a Score question with three ordered levels against the same state, and confirm you get all three typed answers in one call at roughly the same speed as one.

  3. 03

    Create an API key at console.typesafe.ai/settings/keys, export it as TYPESAFE_API_KEY, install typesafe-sdk or @typesafe-ai/sdk, and reproduce the same request in code with client.system_one(state=..., questions={...}).

  4. 04

    Log the model field, every probability and every confidence value from the response, then pick a conservative confidence threshold and route anything below it to a human before you let the answer trigger an action.

  • State the outcome before the background.
  • Provide the real source material.
  • Review the result before expanding the task.
03

Quality control

Check the quality of TypeSafe AI output

Establish that a model's output is reliable enough for the job you are handing it.

  1. 01

    Restate the task, the acceptance criteria, and the failure cost before evaluating any response.

  2. 02

    Test with a fixed evaluation set rather than ad-hoc prompts, so changes are measurable across model versions.

  3. 03

    Check determinism and drift: record model version, parameters, and date, because providers update models underneath you.

  4. 04

    Probe the failure modes deliberately — ambiguous inputs, adversarial phrasing, and out-of-scope requests.

  5. 05

    Validate structure as well as content: confirm schemas, types, and required fields survive real inputs.

  • Pin and record the model version with every result.
  • Evaluate against a fixed set, not one-off prompts.
  • Validate output structure before it reaches downstream code.
04

Privacy & permissions

Use TypeSafe AI safely

Understand what leaves your system on every call, and what the provider keeps.

  1. 01

    Classify the data in each request before it is sent to a hosted model.

  2. 02

    Check the provider's retention and training terms for your specific plan, not the marketing page.

  3. 03

    Store API keys in a secret manager, scope them per project, and rotate them on a schedule.

  4. 04

    Strip or tokenise personal and regulated data before it reaches the request body.

  5. 05

    Set rate limits, spend caps, and monitoring so a runaway loop is bounded and visible.

  • Jev is calibrated, not infallible. "Zero hallucinations" means the answer can never fall outside the options you defined; it can still be wrong within them. Keep arithmetic, counting, and date comparison in code, filter state down to what the question needs, and treat all state as untrusted data because adversarial text can steer an answer.
  • Follow your organisation's approved-use policy.
  • Never treat fluent output as authorization to act.
05

Models & pricing

What TypeSafe AI costs, model by model.

TypeSafe AI sells one model, Jev 1.13, through one endpoint. It is billed on input tokens only — output tokens are free — at $0.042 per million, which the company rounds to "$42 per billion". There are no subscription plans; access is by API key from the early-access waitlist.

TypeSafe AI models, aliases and pricing
Model / aliasResolves toInputOutputRate limits
Jev 1.13 (jev-1.13.0)—$0.042$0250,000 tokens/s · 1,200 requests/min
jev-latestjev-1.13.0$0.042$0Same as underlying model
jev-previewjev-1.13.0$0.042$0Same as underlying model
Prices are USD per million input tokens. Output tokens are not billed. Both aliases currently resolve to jev-1.13.0; the response's model field reports the versioned ID that answered, so pin the version once you have tuned confidence thresholds against it.
What a Jev call actually costs at published rates
WorkloadInput per callCost per callCost per million calls
Short ticket, 3 questions (docs example)312 tokens$0.0000131$13.10
1,000-token record, 10 questions~1,200 tokens$0.0000504$50.40
Maximum context, single request64,000 tokens$0.00269$2,688
Doom demo, 10 queries/second——~$7 per hour (TypeSafe's own figure)
Worked examples at $0.042 per million input tokens. Adding questions to a call costs only their own tokens, so the cheapest design packs every question that shares a state into one request. Rate limits are adjusting dynamically during early access and can change without notice.

Figures verified 17 September 2026 against TypeSafe docs: Models, pricing and rate limits ↗. Model pricing changes frequently — confirm before you budget.

Frontier and default-tier pricing across the three major providers
ProviderFrontier tierBalanced defaultLow-cost tier
OpenAIGPT-5.6 Sol — $5 / $30GPT-5.6 Terra — $2 / $12GPT-5.6 Luna — $0.20 / $1.20
AnthropicFable 5.1 — $10 / $50Sonnet 5 — $2 / $10Haiku 4.5 — $1 / $5
GoogleGemini 3.1 Pro — $2 / $12Gemini 3.1 Pro — $2 / $12Gemini 3.5 Flash-Lite — $0.15 / $1.25
Input / output USD per million tokens, standard context, before batch or caching discounts. Compare tiers by role rather than by name — vendors label their tiers differently.

Cross-provider comparison verified 2 September 2026. Sources: OpenAI API pricing ↗ · Claude platform pricing ↗ · Gemini API pricing ↗

06

Common questions

TypeSafe AI questions people actually ask.

How much does TypeSafe AI cost?

As of 17 September 2026, $0.042 per million input tokens ($42 per billion) on Jev 1.13, with output tokens free. There are no plans or seats: you get an API key and pay per token. A typical short-ticket call with three questions is about 312 tokens, or roughly a hundredth of a cent. TypeSafe says it expects prices to fall rather than rise, and that higher rate limits are available on custom and enterprise plans via sales@typesafe.ai.

What is a System One Model?

A class of model TypeSafe introduced on 15 September 2026, named after Daniel Kahneman's fast, intuitive System 1 thinking. Instead of generating text one token at a time, it takes a state (text or JSON) plus typed questions and returns all the answers in parallel as probabilities: a Choice over options, a Score over ordered levels, or a Noul (yes/no probability). Jev is the first public System One Model.

Is Jev just a small LLM with structured outputs?

TypeSafe says no: Jev uses a different architecture, a parallel sampler rather than autoregressive decoding, and a different training objective (Reinforcement Learning for Calibrated Decisions, RLCD) instead of RLHF or RLVR. The practical difference from an LLM's JSON mode is that Jev cannot produce anything outside your defined options, returns a calibrated probability for every option in one pass, and answers in 70–500 ms.

Does TypeSafe AI really have zero hallucinations?

In one narrow sense, yes: the output is constrained to the options you supplied, so a type error or an invented tool call is impossible by construction. In the everyday sense, no: Jev can still choose the wrong option, and its own documentation lists eight known failure modes for Jev 1.13, including literal reading, counting, date comparison and susceptibility to adversarial text. The calibrated confidence score is what makes the wrong answers manageable — you route low-confidence cases to a person.

How fast is Jev compared to GPT or Claude?

TypeSafe quotes 70–500 ms end to end, from its US West Coast service. On its own four published workflow evals, Jev averaged 0.4 seconds and $0.0004 per case at 67.8% agreement with the reference labels, against GPT-5.6 Terra at 10.1 s / $0.0304 / 67.9% and Claude Opus 5 at 37.8 s / $0.1761 / 73.1%. The 193.6× faster and 444.6× cheaper headline figures come from those evals and TypeSafe itself calls them the higher end of real-world gains.

What can't Jev do?

It does not generate text, code or explanations, and it accepts text only (no images, audio or video yet). Choice questions support up to 255 options. Context is 64k tokens for state plus all questions, and 32k for the state plus the longest single question. It is weak at arithmetic, counting, numeric comparison and date ordering, and it reads instructions literally. TypeSafe's guidance is to keep those in code and give the model only the judgment.

Who is behind TypeSafe AI?

A San Francisco lab that emerged from stealth on 15 September 2026 with $40 million in seed funding led by DCVC. The founders are Diogo Almeida (CEO, former OpenAI and Google Brain researcher and a co-inventor of RLHF and InstructGPT), Sasha Sheng (COO, ex-Meta FAIR) and Erik Gafni (CTO, repeat founder, early at Invitae and Freenome).

07

Core workflows

Step-by-step ways to use TypeSafe AI for the work it does best.

Each workflow is a separate guide with its own steps and review checkpoints.

01 WorkflowRoute a support ticket with confidence-gated actions in TypeSafe AI

Ask Jev three typed questions about one ticket, then let code decide what runs automatically and what waits for a person.

 02 WorkflowGuardrail an LLM's output with a Noul check in TypeSafe AI

Put a sub-second, sub-cent yes/no check between a generative model and the action it wants to take.

 03 WorkflowClassify a large corpus with speculative fan-out in TypeSafe AI

Pack every question a record might need into one call and let code discard what it does not use.

08

Official references

Check the current product documentation.

Features, plan limits, availability, and data controls change. These official pages are the starting points used for this guide.

Explore the Troiana AI Hub →