Anthropic decision model
Searches for an "Anthropic decision model" are climbing as Google and OpenAI ship decision-focused models — but as of 21 November 2026 Anthropic has not announced one. Claude is a generative LLM: you can coax decision-shaped output out of it with tool use and structured outputs, but that's prompting a chat model and parsing the reply, not a model built to return one typed, calibrated answer. Here's what the category is, who's actually in it, and how a purpose-built decision model like Jev differs from asking Claude.
A "decision model" is a model whose entire job is to turn context plus a finite answer space into one typed answer your code can act on — a label from a fixed set, a bounded score, or a calibrated yes/no — instead of free-form prose you then have to parse. TypeSafe AI defined that shape with Jev in September 2026; OpenAI followed with its Decisions API in October, and Google launched Gemini Decision 1 on 11 November 2026. Anthropic, so far, has not shipped anything in this category.
Does Anthropic have a decision model or a decisions API?
No — not as of 21 November 2026. Anthropic's public lineup is Claude, a family of generative large language models (Haiku, Sonnet, Opus, and the newer Fable/Mythos tiers), available through the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. There is no separate "Anthropic Decisions API," no decision-only endpoint, and no decision-specific model announced on Anthropic's newsroom or in its API docs. If that changes it would appear on Anthropic's official channels first — this page is about what exists today, not speculation about a product that hasn't been announced.
How to get decision-shaped output from Claude today
You don't need a dedicated decision model to get a bounded answer out of Claude — you emulate one. Two mechanisms do most of the work:
- Tool use (function calling): define a tool whose JSON schema is your answer shape — an enum for the label, a number with min/max for a score — and force the model to call it. Claude returns arguments that fit the schema, so the label stays in-set.
- Structured outputs: constrain the response to a JSON schema directly, so the fields and types are guaranteed even without a tool call.
Both keep the shape honest, and they're the right move when you're already running Claude for a task and want one more bounded answer out of it. What neither gives you is the thing a decision model is built around: a single, fast, non-generative pass, and a confidence number trained to track real outcomes. Claude still decodes the answer token by token as a side effect of being a language model, and any "confidence" you read off it is not calibrated — it wasn't trained so that 0.8 means right about 80% of the time. For a one-off or low-volume check that's fine; for a gate you call thousands of times in an agent loop, the latency and the untrustworthy threshold start to matter.
Decision model vs asking Claude
| Claude (tool use / structured output) | Jev (decision model) | |
|---|---|---|
| What it is | Generative LLM constrained to a schema | Purpose-built System One decision model |
| How the answer is produced | Decoded token by token, then parsed | One non-autoregressive pass, typed value out |
| Answer stays in-set | Yes, via enum schema / forced tool | Yes, by construction (choice · score · noul) |
| Confidence | Uncalibrated — not trained to match outcomes | RLCD-calibrated 0–1 you can threshold on |
| Latency for one decision | Full generation round trip | ~70–500ms, 40–200x faster than a frontier LLM |
| Best for | A bounded answer alongside work Claude already does | High-frequency routing, grading, gating in agent loops |
The honest read: if you're already in Claude and need one structured answer, tool use is the pragmatic path and you don't need anything else. The case for a decision model shows up when the decision is the hot path — called constantly, gated on a threshold, and sensitive to latency — because then you want a model that emits a typed value in a single pass and a confidence number trained (via RLCD, Reinforcement Learning for Calibrated Decisions) to actually mean what it says. That's what Jev is for, and it's generally available today.
Prefer Claude for the heavy lifting? Use both.
This isn't either/or. A common pattern is to keep Claude for generation and reasoning, and offload the small bounded calls — "is this in scope," "which of these five tools," "does this need a human" — to a decision model so they're fast and cheap. The decide call takes a finite answer space in and returns one typed answer plus a calibrated confidence out:
// Offload a bounded decision; keep Claude for the prose
const res = await fetch("https://jevtypesafeai.com/api/v1/decide", {
method: "POST",
headers: {
"Content-Type": "application/json",
Authorization: `Bearer ${process.env.JEV_KEY}`, // jv_live_...
},
body: JSON.stringify({
state: { message: text },
questions: {
route: { type: "choice", instructions: "Where should this go?",
criteria: { answer: "a factual question", escalate: "needs a human", ignore: "spam or noise" } },
urgent: { type: "noul", instructions: "Does this need attention within the hour?" },
},
}),
});
const { answers } = await res.json();
if (answers.urgent.noul > 0.8) escalate(); // calibrated threshold
else handle(answers.route.choice); // always in-setOutput tokens are free and a single decision costs a fraction of a cent; the exact per-million input figures are on the pricing page. Official upstream is POST https://api.typesafe.ai/v1/systemone.
FAQ
Does Anthropic have a decision model or decisions API?
No. As of 21 November 2026 Anthropic has not announced a dedicated decision model or a decisions API. Claude is a family of generative LLMs available through the Claude API, Bedrock, Google Cloud and Microsoft Foundry. OpenAI (Decisions API, Oct 2026) and Google (Gemini Decision 1, Nov 2026) have shipped decision models; Anthropic has not.
Can I get a decision-shaped answer out of Claude?
Yes — use tool use (force a tool whose JSON schema is your answer shape, e.g. an enum) or structured outputs to constrain the reply to a schema. The label stays in-set, but Claude still decodes it token by token and any confidence it reports is not calibrated. For a low-volume check that's fine; for a high-frequency gate, a purpose-built decision model is faster and gives a threshold you can trust.
What's the difference between a decision model and prompting Claude?
A decision model (like Jev) returns a typed value in a single non-autoregressive pass with an RLCD-calibrated confidence — built and trained only for decisions. Prompting Claude with a constrained schema gives you the right shape but through a generative LLM: token-by-token decoding and uncalibrated confidence. Same answer shape, different machinery and different cost/latency profile at scale.
Should I use Claude or a decision model?
Use Claude when you need generation or reasoning and want one bounded answer alongside it. Reach for a decision model when the decision is the hot path — routing, grading, or gating called thousands of times in an agent loop, where latency and a calibrated threshold matter. Many apps use both: Claude for the prose, a decision model for the fast bounded calls.
Will Anthropic release a decision model?
There's no public announcement either way as of November 2026. If Anthropic ships one it would appear on its official newsroom and API docs first. Until then, the way to get calibrated, typed decisions is either to constrain Claude with tool use / structured outputs or to use a purpose-built decision model such as Jev.
See also: Jev vs Claude · What is a decision model · OpenAI Decisions API explained · Gemini Decisions API explained · Microsoft-Decision-1 explained · Playground
Need a decision, not an essay? Try one free.
Anthropic hasn't shipped a decision model — Jev is one, and it's generally available. Run a real typed, calibrated decision in the browser with no signup, then grab a jv_live_ key.