Fast Typed AI Models and Chat Models: When to Use Each

Ask a typed model a question and the answer arrives in the shape you asked for: one option from a list you wrote, a position on a scale you described, or the probability that the answer is yes. Your code reads that directly and moves on. Ask a chat model the same question and you get sentences, which a person reads and which your code has to pull apart before it can act. What Jev returns and what it costs covers the typed side in detail.

The second difference is how much you can put in front of the model at once. TypeSafe publishes the typed model's context as “64k tokens per request; 32k tokens for state plus the longest question,” with text only input on the model page. Anthropic's model overview lists a context window of 1M tokens for its larger current models and 200K tokens for Claude Haiku 4.5. When a job needs the model to hold a long document and reason across the whole thing, that gap settles the question for you.

Four shapes cover most of it. Each one has a pile of items and a single judgment you could teach a new hire in two sentences.

• Sorting. A year of form submissions dropped into service categories, or reviews grouped by the complaint behind them.
• Scoring. Every page, listing or ad rated against a scale you wrote yourself, so the whole set carries a number you can sort on.
• Routing. New inquiries sent to the right inbox, crew or priority queue, which pairs with the handoff rules in lead triage automation.
• Checking. Whether every item in a batch contains the thing you said it would, run as one yes or no question per rule.

Volume is what makes this worth wiring up. Under a hundred items a month you'll do it faster by hand. At a few thousand a month, the typed pass finishes while you're still deciding where to start, and TypeSafe publishes throughput of “250,000 tokens per second / 1,200 requests per minute” on the same model page.

The other half is anything where the output is the words themselves. That covers rewriting a service page so a homeowner can tell what you do and where you do it, drafting a reply to a customer who's unhappy, writing the script that crawls your site and batches the requests, and reading a table of results to explain the pattern in it and argue through what you should do next.

A typed model can't take those on, and its own documentation says so: the page on where Jev falls down describes the model as “not trained to generate text.” That's a design choice, and it's why the two kinds of model sit next to each other in a working process.

The two bill differently, and that shapes which one you point at a big pile. TypeSafe's model page prints the typed model at $42 per Btok and $0.042 per Mtok, where a Btok is a billion tokens and an Mtok is a million, and adds that “Output tokens are free,” so you pay for what you send. Anthropic's pricing page shows a base input tokens column and an output tokens column, with output priced higher on every current model: Claude Haiku 4.5 at $1 per MTok in and $5 per MTok out, and Claude Sonnet 5 at $2 and $10. Both pages move, so read them yourself before you build a budget on them.

Neither vendor prints a latency figure on the pages I checked, so measure that on your own work. What the docs do show is the size of the answer, and TypeSafe's documented example response reports 20 output tokens for a single yes or no question. TypeSafe also notes that its published rate limits “can change without notice” while it serves current demand.

Here's the arithmetic on a two step job, and every token count in it is my assumption while the prices come from those two pages. Say you have 2,000 pages and you ask four typed questions about each one in a single request, and say each request works out to roughly 1,500 input tokens. That's 3 million input tokens, which at $0.042 per million comes to about 13 cents for the whole pass. Say 60 of those pages come back weak and you hand each one to Claude Haiku 4.5 for a rewrite at roughly 2,000 input tokens and 1,200 output tokens. That's 120,000 input tokens at about 12 cents plus 72,000 output tokens at about 36 cents, so near enough 48 cents for the writing. The scoring pass runs to a fraction of a cent per page and the rewrite runs to close to a cent per page, which is the whole reason you score first and write second.

The pattern that works puts them in order. Gather the items in code, ask the typed model the same short questions about every one, sort by the answers, then hand only the worst cases to a chat model to fix. The typed pass tells you where to look, and the chat pass does the work. Using Jev with Claude to fix your website walks the whole loop, and deciding which fixes come first covers what to do with the ranked list once you have it.

Two guardrails come straight from TypeSafe's page on the model's weak spots. Keep arithmetic and date comparison in your own code, since the page says “Jev is not a calculator” and describes the model reading dates as plain text. And filter the request before you send it: “Accuracy falls as the state grows with content unrelated to the decision.”

This is the rule that keeps both models useful over a long stretch. Page copy, review replies, ad text, email and anything posted to a listing all get a person's eyes before they go live. The models do the sorting and the drafting, and a human being owns what goes out under your name.

Google's guidance on helpful content is direct about the risk of skipping that step. It says the reason a page exists should be “that you're creating content primarily to help people, content that is useful to visitors if they come to your site directly,” and that using automation, including AI generation, to produce content “for the primary purpose of manipulating search rankings” is “a violation of our spam policies.” It also asks publishers to check their own work: “Evaluating your own content against these questions can help you gauge if the content you're making is helpful and reliable.”

A typed model makes that rule easier to keep, because System One models are trained so their probabilities are “optimized against outcomes to reflect uncertainty.” You can gate on the number: send every low confidence answer to a person, and set a higher bar for anything hard to undo. Human review for AI marketing and the human review playbook both lay out where in a process the judgment belongs, and AI SEO and automated content production is worth reading before you point any model at a content calendar.

1. Write down the output you need. If a spreadsheet column would hold it, a typed model fits. If a paragraph is the output, a chat model fits.
2. Count the items. Dozens a month means you should do it by hand or in a chat window. Thousands a month pays back the effort of wiring up a typed pass.
3. Name the cost of a wrong answer. A misrouted inquiry costs a callback. A published page with a wrong claim in it costs more, so that one waits for a person.
4. Move the arithmetic into code. Dates, totals, thresholds and deduping belong in a script you can test.
5. Decide who signs off, by name. An approval step with nobody's name on it gets skipped by week three.

Start with one decision you already make by hand every week and run fifty items through both approaches before you commit to either. If you'd rather see a ranked list of your own pages before you build any of this, our Jev Website Scanner, which runs Jev on your site, is the quickest way to get one. The first scan of a site is free for up to 100 pages, and you enter an email so we can send you the report link.