Comparison

Jev vs LLMs

Jev vs LLMs in 2026 — when TypeSafe's decision model can replace ChatGPT, Claude or a cheap small model, what it costs per 1,000 decisions against each, and what it cannot do.

Updated September 2026

Quick answer: Jev costs less per decision than the LLMs it is compared with, but the size of the saving depends on what it replaces. On TypeSafe’s own evaluation, Jev costs about 8 times less per case than OpenAI’s cheapest GPT-5.6 model, Luna, at almost the same accuracy (67.8% against 66.8%), and about 440 times less than Claude Opus 5. At list prices with reasoning switched off, the gap to the cheapest small models narrows to 2 to 4 times. Jev does not write text: it picks from options you define, scores against criteria you set, or returns the probability that the answer is yes. Use Jev for high-volume decisions with a fixed set of answers, and an LLM for anything that has to write, calculate or reason about dates.

At a glance

What Jev does that an LLM does not

An LLM answers a routing question by writing text that your code then parses. Jev answers it in the type you asked for. Given a support ticket and a list of teams, Jev returns the team it picks, the probability of each team and a confidence figure, rather than a sentence naming one (TypeSafe docs).

Jev answers three kinds of question, which TypeSafe calls primitives:

• Choice picks from up to 255 options you define.
• Score rates the input against criteria you set.
• Noul returns the probability, from 0 to 1, that the answer to a yes/no question is yes.

TypeSafe trains Jev to give honest probabilities rather than answers people prefer, and says its confidence figures are calibrated: across many predictions, higher confidence means higher accuracy. It produces its answers in a single pass rather than one token at a time, which TypeSafe credits for its speed (TypeSafe). The Jev model page covers how it works in full.

What it costs against each kind of LLM

TypeSafe’s evaluation

TypeSafe publishes the one head-to-head that includes cheap models. It ran four decision workflows (security incidents, agent trace review, invoice processing and customer service) through each model and scored each against a reference answer: the average of GPT-6 Astra and Claude Fable 5.1 at high reasoning. Every other model ran at its provider’s default reasoning setting (TypeSafe evals).

Jev matches the agreement of the cheap and mid-priced models in the table and trails GPT-5.6 Sol and Claude Opus 5 by 6.3 and 5.3 points. The multiples use TypeSafe’s rounded $0.0004 figure for Jev, so they are approximate.

Four limits apply to these figures. They come from workflows TypeSafe built itself. Agreement with two frontier models is not the same as accuracy against answers checked by people, and TypeSafe itself says some bias could exist. The LLMs ran with reasoning switched on by default, which adds output tokens and time. And the table predates GPT-6 Luna (22 September 2026), which costs half GPT-5.6 Luna’s input price. SiliconANGLE notes that TypeSafe’s performance claims have not been independently verified (SiliconANGLE).

At list prices, with reasoning off

A small LLM with reasoning switched off writes only a few tokens per decision, so its cost comes almost entirely from input, and the comparison with Jev comes down to input price. The table below assumes each decision reads 1,000 input tokens and an LLM writes 10 output tokens back, at each provider’s list price on 30 September 2026. Jev’s output is free.

At one million decisions a month on these assumptions, Jev costs about $42, GPT-6 Luna about $105 and Claude Opus 5.5 about $4,200. Prices are from TypeSafe, DeepSeek, Anthropic and the linked model pages.

Three things move these numbers:

• Reasoning. An LLM that reasons before answering bills for those tokens as output, which is why the per-case costs in TypeSafe’s evaluation are several times higher than list-price arithmetic suggests.
• Caching and batching. LLM providers discount repeated prompt text: Anthropic charges 10% of the input price for cached tokens on Claude Haiku 4.5, and DeepSeek $0.003 per million off-peak. Anthropic’s Batch API halves prices for work that can wait.
• Tokenisers. The same text produces different token counts on different models. Anthropic says its models from Claude 4.7 onward produce about 30% more tokens for the same text than earlier ones.

No published measurement covers the accuracy of a small LLM with reasoning switched off on the same tasks, so the list-price table compares cost only.

When Jev fits and when an LLM does

Jobs that suit Jev have a high volume of repeated decisions and a fixed set of possible answers. TypeSafe’s cookbooks cover these (TypeSafe docs):

• Routing requests or tickets to the right team or intent
• Classifying documents and retrieval passages, including hierarchical categories
• Re-ranking search results
• Checking citations, and guardrail checks on LLM output
• Choosing which function or tool to call
• The four workflows in TypeSafe’s evaluation: security incident triage, deciding whether an agent’s trace needs human review, invoice processing and customer service

TypeSafe’s own patterns pair Jev with LLMs rather than replacing them. One uses Jev’s confidence figure to decide when to pass a decision to a reasoning model or a person. OpenRouter also lists Jev Router, which uses Jev to choose a model and reasoning effort for each request.

Jobs that need an LLM:

• Anything that produces text, code, summaries or explanations
• Arithmetic, counting and comparing numbers
• Reasoning about dates and times
• Open-ended questions without a fixed set of answers, or more than 255 options
• Images, audio or video input
• Inputs longer than Jev’s context: 32k tokens for the input and longest question

Limits

TypeSafe publishes a list of where Jev 1.13 goes wrong (TypeSafe docs):

• Maths and numbers. TypeSafe says “Jev is not a calculator”. It does not count reliably or judge whether two values are close.
• Dates and times. It reads dates as text, not as ordered quantities.
• Literal reading. It answers the question as written, not as intended.
• Indirection. Double negatives and nested instructions are answered less reliably.
• Irrelevant input. Accuracy falls as the input fills with content unrelated to the decision.
• Adversarial content. Text written to steer the model can move the answer.
• Consistency. Related questions are not guaranteed to produce consistent answers.

On TypeSafe’s evaluation, Jev agrees with the frontier reference on 67.8% of cases, 6.3 points below GPT-5.6 Sol and 5.3 below Claude Opus 5 (TypeSafe evals).

How to get access

Jev is available from four places, all at $0.042 per million input tokens with free output:

TypeSafe opened early access at launch on 15 September 2026, admitting developers from its waitlist. Vercel added Jev to its AI Gateway on 16 September, and says it adds no markup to token prices (Vercel). Cloudflare lists it as generally available, with cached input free. More detail is on the Jev model page.

Choose Jev if…

• You make the same kind of decision thousands or millions of times, and the possible answers are known in advance.
• You currently parse labels out of an LLM’s text, and want the answer returned in a fixed type with a probability attached.
• Response time matters: TypeSafe reports 70 to 500 milliseconds end to end.
• You want a confidence figure to decide when to escalate a decision to a larger model or a person.

Choose an LLM if…

• The job produces text, code or explanations.
• The decision involves arithmetic, counting or dates.
• The answers are open-ended, or the input is longer than 32k tokens or includes images, audio or video.
• You need the highest agreement with frontier models: on TypeSafe’s evaluation, GPT-5.6 Sol and Claude Opus 5 scored 6.3 and 5.3 points above Jev.

Frequently asked questions

Is Jev a replacement for ChatGPT or Claude?

No. Jev cannot write text, code or explanations. It returns typed decisions, such as a choice from a list, a score or a yes/no probability, so it can replace an LLM only for decisions with a fixed set of answers. TypeSafe’s own patterns pair Jev with an LLM rather than replacing one.

Is Jev cheaper than ChatGPT?

Yes, per decision. On TypeSafe’s evaluation, Jev cost about $0.0004 per case against $0.0033 for GPT-5.6 Luna, $0.0304 for GPT-5.6 Terra and $0.0836 for GPT-5.6 Sol. At list prices with reasoning switched off, GPT-6 Luna costs about 2.5 times as much as Jev per decision and GPT-6 Astra about 250 times as much.

Is Jev cheaper than Claude?

Yes, per decision. On TypeSafe’s evaluation, Jev cost about $0.0004 per case against $0.0195 for Claude Haiku 4.5, $0.1174 for Claude Sonnet 5 and $0.1761 for Claude Opus 5. At list prices, Jev’s $0.042 per million input tokens is 238 times below Claude Fable 5.1’s $10.

Is Jev as accurate as GPT or Claude?

On TypeSafe’s evaluation, Jev agreed with the frontier reference answer on 67.8% of cases, level with GPT-5.6 Terra (67.9%) and Claude Sonnet 5 (67.8%), above Claude Haiku 4.5 (53.6%), and below GPT-5.6 Sol (74.1%) and Claude Opus 5 (73.1%). TypeSafe built the evaluation, and its figures have not been independently verified.

Is Jev free?

No. Jev costs $0.042 per million input tokens, and its output tokens are free. The price is the same on TypeSafe’s API, OpenRouter, Vercel AI Gateway and Cloudflare Workers AI.

Does Jev work with Claude Code?

Not as a coding model, because Jev cannot write code. TypeSafe publishes an official agent skill that gives Claude Code and other coding agents the context to build applications with the Jev API, installed with claude plugin marketplace add typesafe-ai/skills (TypeSafe docs).

Can I use Jev without the waitlist?

Yes. OpenRouter, Vercel AI Gateway and Cloudflare Workers AI all serve Jev without a TypeSafe account, with a 32k-token context. Calling TypeSafe’s own API, which allows 64k tokens per request, requires early access from TypeSafe.

What is Jev’s context window?

Jev accepts 64k tokens per request through TypeSafe’s API, of which the input plus the longest question can use 32k. OpenRouter, Vercel AI Gateway and Cloudflare Workers AI list a 32k-token context.