AI

AI decision models choose from a list instead of writing. When do they beat chatbots?

Google, OpenAI and Microsoft now sell models that score a fixed set of answers, and the evidence for their advantages comes mostly from the vendors.

12 min read
Black-and-white close-up of tall, tilted stacks of envelopes and paper sheets with their edges fanned out in rows.
Image by cocoparisienne from Pixabay

A decision model takes a question and a list of allowed answers, then returns a probability for each option instead of writing text. Google's Nov. 11 launch joined a field that already included TypeSafe, OpenAI and Microsoft. The design suits narrow, high-volume tasks such as routing, but most performance claims are vendor-reported and untested by outsiders.

A decision model is an AI system that receives a question and a list of permitted answers, then returns a probability for each answer instead of composing a sentence. Google released one, Gemini Decision 1, on Wednesday, Nov. 11, 2026. It joined products from OpenAI, Microsoft and a San Francisco start-up called TypeSafe AI, all of which appeared in the previous two months.

The category matters because many software jobs need a label, not an essay. Sorting a support ticket, checking a message against a policy, or choosing which tool an AI agent should call all end in a short, predictable answer that other code can act on. Vendors argue that a model built only for that step is faster and cheaper than a chatbot, and easier to plug into a workflow.

The evidence behind those arguments is thinner than the marketing. Nearly every performance figure comes from the company selling the model, no peer-reviewed study or independent evaluation of Google's model turned up, and the reporting used here is secondary coverage rather than the companies' own documentation.

How a decision model differs from a chatbot

A chatbot is built on a large language model, or LLM, which writes its answer one token at a time. Each piece of text depends on the pieces before it, so the process has to loop until the answer is complete. According to the analysis site i-scoop, Gemini Decision 1 works differently: it reads the input once and scores every allowed option in a single forward pass.

Trade outlet ITdaily, summarizing Google's documentation, reports that the model produces probabilities and confidence scores and does not generate answer text. The result is a structured output, such as "72% option B," that a program can read directly without parsing a paragraph.

The three question types in Google's interface

A CASRAI summary of Google's model documentation describes three formats. A yes/no question returns a probability of "yes" between 0.01 and 0.99, along with a confidence value. A choice question evaluates between 2 and 255 options and returns the selected option, a probability for each option, and a confidence value. A score question grades the input against a rubric of 2 to 255 ordered levels.

Per the same summary, one request can carry up to 1,000 independent questions about a shared input. The questions cannot see one another's answers, so the calling code has to combine the results. The model accepts text, images, audio, video and PDFs, with a context of 262,144 tokens shared across the input and all questions.

The billing follows from that structure. One secondary source reports that the shared input is billed once per request regardless of how many questions are asked, and each question adds only its own instruction and criteria tokens. That detail rests on a single source and was not checked against Google's pages.

A crowded field, not a new entrant

Google's launch was not the third entry in the category. Per coverage relayed by Medium and TrueFoundry, TypeSafe AI came out of stealth with a model called Jev on Sept. 15, 2026. how2shout reports that the company was founded by former OpenAI researcher Diogo Almeida. The Medium post and other sources agree on the date.

OpenAI's timeline is less clean. CryptoBriefing reports a limited preview of its Decisions API at its DevDay event on Sept. 29, Unite.AI reports a public beta on Oct. 6, and CryptoBriefing reports general availability on Oct. 22. Microsoft's Decision-1 arrived on Oct. 9, according to Seeking Alpha, so it landed after OpenAI's beta and before OpenAI's general release. Who led whom depends on which milestone counts.

Launch timeline of decision models, September to November 2026
Date (2026) Product Company Basis for the date
Sept. 15 Jev (early access) TypeSafe TrueFoundry
Sept. 29 Decisions API (limited preview) OpenAI CryptoBriefing, single source
Oct. 1 Clef, Clef-flash Cloudflare rpabotsworld, single source
Oct. 6 Decisions API (public beta) OpenAI Unite.AI
Oct. 7 d1-3B, d1-omni-600M Liquid AI rpabotsworld, single source
Oct. 9 Microsoft-Decision-1 Microsoft Seeking Alpha
Oct. 22 Decisions API (generally available) OpenAI CryptoBriefing
Nov. 11 Gemini Decision 1 / Decisions API Google CryptoBriefing

The dates for TypeSafe, OpenAI, Microsoft and Google are consistent across several outlets. The Cloudflare and Liquid AI dates come from one blog, rpabotsworld, and were not independently confirmed. TechCrunch also mentions a Liquid model called D1 Ultra released shortly before Oct. 22, which may be a different product from the ones in the table.

Same label, different construction

The products are not built the same way. Per Startup Fortune and other coverage of Microsoft's announcement, Microsoft's model is a post-trained version of Qwen3.5-9B, a 9-billion-parameter open-weight model from Alibaba. TechCrunch, quoting an OpenAI product lead, reports that OpenAI did not train a specially calibrated decision model. The team reconfigured its existing Luna model to run faster and cost less. This is a single-source account of OpenAI's approach.

TypeSafe describes its own training method as Reinforcement Learning for Calibrated Decisions, a label for which no independent paper was found. The aisuperior blog, relaying TechCrunch's original coverage, reports that Jev is a transformer-based model that is "not a large language model." The internal design of Google's model is not described in any of the coverage reviewed, so claims beyond "single pass, no text output" would be guesswork.

Reported prices are similar for the larger vendors. According to coverage of each launch, Google's Decision 1 and OpenAI's Decisions API each cost $0.10 per million input tokens, and Microsoft's Decision-1 costs $0.042 per million input tokens. All three reportedly charge nothing for output, since the output is only scores. Reported limits differ: OpenAI's API takes up to 400 questions per request and handles text and images, while Google's takes up to 1,000 and adds audio, video and PDFs.

What "calibrated" means, and why vendors lead with it

Press coverage defines a calibrated model as one whose stated confidence tracks how often it is actually right. A model that reports 90% confidence should be correct roughly nine times out of ten. This brief did not locate peer-reviewed literature on calibration, so the definition here is the one used in trade coverage.

The pitch is practical. how2shout, paraphrasing TypeSafe's argument, says a model that is right 95% of the time but cannot indicate which 5% it doubts cannot be left to run unattended. If the confidence numbers are trustworthy, software can act on clear cases automatically and send ambiguous ones to a person or a larger model. That routing pattern appears repeatedly across the coverage.

Google's reported numbers

Google's comparison, as reported by SoftTechHub and matched for the first three figures by Aiera and MarkTechPost, uses top-1 expected calibration error, where lower is better. Gemini Decision 1 scored 0.0250, TypeSafe's Jev 1.14 scored 0.0494, OpenAI's GPT-6 Luna scored 0.0721, and Gemini Flash-Lite, a chatbot-style model, scored 0.2104. On that measure Google's model has about half the error of Jev's.

Several caveats apply. The figures come from Google's own benchmark, and the dataset and sample size were not found in the coverage, so whether the gaps are statistically meaningful is unknown. The same reported comparison gives accuracy on a benchmark named JevBench of 92.2% for Google, 90.9% for OpenAI's model and 90.0% for Jev. The origin of JevBench could not be verified, and its name suggests a link to a competitor, so whether it favors any vendor is unclear.

Independent measures paint a less tidy picture. TechCrunch reported on Oct. 22 that third-party indices, including the AIM System One Index and the Artificial Analysis Decision Index, tended to favor Liquid's D1 Ultra and Base Lab's Elephant-1. On those two indices Jev ranked seventh and third, and OpenAI's model ranked eighth and fourteenth. Gigazine relayed a Google post citing a score of 66 on version 2.1 of the Artificial Analysis index, but Google's rank was not found.

Where decision models win

The strongest case is a narrow, high-volume task with a closed set of answers. Google names tool selection, policy checks, model routing and context grading as target uses. In each, the output is a label that code consumes, and a paragraph of explanation would only add cost and delay.

Speed

Skipping the token-by-token loop makes a speedup plausible from the design alone. The size of the gain is a vendor claim. Per Apidog, Google reports a median of 50 to 100 milliseconds of model compute and 150 to 250 milliseconds end to end, with latency nearly flat from 1 to 100 questions for inputs up to 16K tokens. Towards AI reports TypeSafe's latency at about 70 to 500 milliseconds and a 20-to-200-times speedup, and a headline cited by rpabotsworld claims 35 times for Microsoft against GPT-6 Sol. These figures use different baselines and cannot be compared with one another.

The New Stack, relaying Google, reports that Google already uses the model as a router in Gemini Notebook, with latency reductions of up to 40%. AlternativeTo, a single secondary source, adds that Google Search uses it to detect adversarial behavior involving Model Context Protocol tools. Both are self-reported by Google.

Cost

In Google's own test, running a shared suite cost $0.88 on Decision 1 against $4.14 on Gemini 3.5 Flash-Lite through the Interactions API. That is roughly a fivefold difference, but it compares one vendor's two products on a suite the vendor chose. It is not an independent price comparison, and the result depends on how a team would otherwise set up a cheap chatbot tier.

Usable output and the "both" architecture

Because the output is a fixed structure, there is nothing to parse and the answer takes one of a known set of values. Developer commentary recommends pairing the two kinds of model rather than choosing between them. One developer's Medium post, which is opinion and not a vendor statement, puts it this way:

The best architecture uses both: Jev decides, the LLM writes, and your code connects them.

A developer's post on Medium

Where they lose, and how they can fail

The clearest limit follows from the design: a decision model cannot write or explain. Coverage of Google's documentation says Gemini Decision 1 does not support text generation, thinking, function calling, search grounding, code execution or computer use. A developer who needs a traceable explanation has to add a language model or document the decision logic elsewhere in the workflow. Learnetto's guide also lists free-form extraction, multi-step reasoning and tool calls with arguments as jobs for a generating model.

The list is a closed world

CryptoSlate, summarizing Google's documentation, reports that the model spreads probability among the supplied options even when none of them fits. A chatbot asked a question outside its categories can say so. A model forced to choose will still produce a ranking, and the output will look confident. Designers have to include an "other" option or a threshold that sends low-confidence cases elsewhere. The summary was not checked against Google's documentation.

Manipulation by the input itself

The same coverage reports a warning in Google's documentation that user input can argue for its own classification, for example by claiming it deserves approval. The recommended defenses are explicit criteria, adversarial testing, and a human or second check before high-stakes actions such as payouts. No independent tests of this weakness were found.

"Can't hallucinate" is only narrowly true

TypeSafe markets its model as one that cannot hallucinate. TrueFoundry grants a narrower version: a model that never emits free text cannot invent a citation or a tool name. It can still choose the wrong option, and its reported confidence can still be miscalibrated. A wrong answer delivered as a clean 0.97 may be harder to catch than a rambling one.

Calibration may not travel

Calibration is measured on benchmark data, and sources including a DEV Community analysis note that it is untested on messy real-world categories and shifting data. Calibration can shift with a new domain, language, class balance or criteria. Google's own guide, as relayed in the coverage, tells developers to set thresholds using labeled examples from their own data instead of trusting published scores.

Product-specific limits also apply. Per the coverage of its documentation, Jev is weak at math, date comparisons and long stretches of irrelevant context, and accepts only text. Google's endpoint reportedly offers no context caching, Batch API or Flex and Priority inference, so identical input sent in two requests is billed twice. It is hosted only by Google, with no open weights or self-hosting, and is in the v1beta stage.

What to watch next

Picking a label from a list is classification, a long-standing machine learning task, and nothing in the sources suggests the task itself is new. What is new is the packaging: large general-purpose models sold as an endpoint that returns only probabilities, priced by input and with free output. Whether that packaging beats a well-tuned small classifier or a cheap chatbot tier is a question the available evidence cannot settle.

The practical split looks fairly stable. Decision models fit narrow, repeated steps where speed, cost and a machine-readable answer matter. Chatbots fit explanation, open-ended work and multi-step reasoning, and the sources describe a pairing in which one decides and the other writes. For anything with real consequences for a person, the vendors' own documentation points to extra checks.

The development to watch is independent testing. A public comparison on held-out data, with calibration measured under distribution shift and with questions that fit none of the options, would show whether "calibrated" describes real behavior or a benchmark result. Until then, the sensible reading is that the design is sound, the vendors' pitches agree with one another, and the size of the benefit is still unproven.

Frequently asked questions

What is an AI decision model?

It is a model that takes a question and a developer-supplied list of allowed answers, then returns a probability for each option instead of writing text. Vendors including Google, OpenAI, Microsoft and TypeSafe sell versions of it. The term is a product-category label, not the older statistics sense of decision trees.

How is a decision model different from a chatbot?

A chatbot generates its reply token by token. According to i-scoop's analysis, Google's model scores all options in a single forward pass and outputs no text. That makes it unable to explain itself, but it returns answers in a structured form that code can use directly.

When is a decision model better than a chatbot?

It is better suited to narrow, high-volume tasks with a closed set of answers, such as routing, moderation, tool choice and rubric scoring. The reasons given are lower latency, lower cost and machine-readable output. The size of the benefit rests on vendor benchmarks.

What does a calibrated probability mean?

A calibrated model's confidence matches how often it is correct, so answers given at 90% confidence should be right about nine times in ten. Vendors lead with this because it lets software act on clear cases and escalate doubtful ones. Google's guide advises testing thresholds on a developer's own labeled data.

Can a decision model hallucinate?

TrueFoundry grants a narrow version of the no-hallucination claim: a model that never writes free text cannot invent a citation or tool name. It can still pick the wrong option or report miscalibrated confidence. It may also spread probability across options even when none fits.

Sources (24)

  1. Google releases Gemini Decisions API for fast multiple-choice decisions , ITdaily
  2. Google's new AI can make 100 decisions in under a second, without writing an answer , CryptoSlate
  3. Gemini Decision 1 turns AI answers into probabilities, and that is what European firms can govern , i-scoop
  4. Gemini Decisions API: Launch and Research Uses , CASRAI
  5. OpenAI Releases Decisions API in Public Beta, Powered by GPT-6 Luna , Unite.AI (2026-10-06)
  6. OpenAI makes its Jev competitor widely available as decision model competition heats up , TechCrunch (2026-10-22)
  7. OpenAI makes its Decisions API generally available at $0.10 per million input tokens , CryptoBriefing (2026-10-22)
  8. Google launches Gemini Decisions API for fast multimodal choices , CryptoBriefing (2026-11-11)
  9. Microsoft's Nadella unveils tech giant's new fast decision-making AI model , Seeking Alpha (2026-10-09)
  10. Microsoft Launches Decision-1, a Model Built to Score Choices, Not Chat , Startup Fortune (2026-10-12)
  11. Microsoft Decision-1 and the Rise of Decision Models (2026) , rpabotsworld
  12. Google releases 'Gemini Decisions 1,' an AI model for decision-making , Gigazine (2026-11-12)
  13. Google launches new Gemini Decisions API for better decision-making in AI applications , AlternativeTo
  14. Google built an AI model that can't write a sentence — and that's the point , The New Stack
  15. Gemini Decisions API: Developer Guide , Learnetto
  16. Google's Decisions API turns Gemini into a control plane, not a chatbot , DEV Community
  17. Gemini Decisions API: What Gemini Decision 1 Does, Pricing and How It Works , Apidog
  18. Gemini Decisions API Explained: Scores, Not Text , Aiera
  19. Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens , MarkTechPost (2026-11-11)
  20. Gemini Decisions API: Google's Model That Writes No Text , SoftTechHub (2026-11-11)
  21. Jev by TypeSafe AI: From "Smart If-Statement" to Production — A Complete, Hands-On Guide , Medium (aeontanvir)
  22. TypeSafe AI's Jev: What "System One Models" Actually Are , TrueFoundry
  23. An Ex-OpenAI Researcher Built an AI Model That Can't Write a Sentence , how2shout
  24. What Is Jev? TypeSafe's Jev AI Model , AI Superior

Share this article

Back to list