Google shipped a Gemini that never writes a sentence, and it changes how you pick a model
Google's Gemini Decision 1 only scores answers you write in advance, and OpenAI already sells the same kind of model, so asking "ChatGPT, Claude or Gemini?" no longer gets you far. This playbook shows how to pick a model for each job, test it on your own cases and price the whole task.
Neo NeumannAI Practice LeadNovember 12, 2026
On 11 November 2026, Google launched a Gemini model that does not write. The Gemini Decisions APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → is built for fast, closed-ended decisions like yes/no questions, multiple choice, and scale-based ratings, and the model behind it, Gemini Decision 1, can take up to 1,000 questions about one input in a single pass, at $0.10 per million input tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → with zero output tokens billed. OpenAI got there first. On October 22 the company made its Decisions API generally available, a dedicated endpoint that asks GPT-6 Luna questions with a fixed set of possible answers and returns a structured result rather than prose. In the same few weeks, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, and Gemini 4 Argon became generally available on October 13, 2026. Ask "ChatGPT, Claude or Gemini?" today and each vendor can answer with a whole range of models, from verdict machines that cost a fraction of a cent to frontier models priced for the hardest work.
The thesis of this playbook: the vendor name is now the least useful thing to decide on. All three vendors sell models built for different shapes of job, and the independent scoreboards often disagree with the launch charts. So you choose one job at a time. Write down what the output looks like, sort the job by shape, test two or three candidates on your own files, and price the full task. Once you have done that, the brand is just a line in the purchase order.
How to choose between ChatGPT, Claude and Gemini for a job?
Choose by the shape of the output first, the stakes of an error second, and the vendor last. In practice that comes down to six moves.
Write the job down as an output spec
Before you open any chat window, write three lines: what the finished output looks like, who acts on it, and what a wrong answer costs. "Tag each incoming contract clause as standard, needs review or reject" and "draft our reply to the client's renegotiation letter" both count as "legal work". They need completely different models. The first has three possible answers you can list in advance. The second needs words nobody has written yet.
Sort the job with the verdict-or-prose rule
Here is the decision rule, short enough for a whiteboard:
If the answer fits in a list you can write in advance, buy a verdict. If it needs words nobody has written yet, buy a generator. If a wrong answer costs more than an hour of someone's time, buy the top tier and keep a human on the check.
The verdict models are a new category, and they behave differently. Rather than generating answer text tokentokenA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → by token, Gemini Decision 1 scores candidate options and returns probabilities, with a confidence score from 0 to 1 on every answer. Google (the vendor, so read this as a sales claim) says it already uses the API internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Google Search. OpenAI took a different route to the same product. Nikunj Handa, who leads product for OpenAI's APIs, told the publication that the company did not train a separate model but reconfigured Luna for faster, cheaper inferenceinferenceThe moment a trained AI model is put to work: it takes a new input and produces an answer, prediction or generated output.View full definition →, cutting responses from one to three seconds to roughly 100 to 180 milliseconds.
The probability is the useful part for a manager. Among decisions assigned a probability of 0.7, a well-calibrated model should be correct about 70% of the time. If calibration holds on your data, you can set a policy: accept answers above a threshold automatically and send everything below it to a person.
Shortlist one tier per vendor
Within the generator category, the vendors themselves tell you to pick by tier. When Anthropic launched Fable 5.2, CNBC reported that other models like Claude Opus 5.5, which Anthropic released in September, are better suited for daily tasks, so Penn said users should turn to Fable 5.2 when they need to tackle hard problems in engineering, mathematics and robotics. Google, with the same commercial interest, says Gemini 4 Argon rolled out globally for complex coding, financial analysis, and legal workflows. OpenAI's GPT-6 line comes in tiers too: GPT-6 Luna itself is a general text model, released 22 September at $0.10 / $0.50 per 1M tokens.
So your shortlist for a prose job is one tier from each vendor that matches the stakes, usually two or three candidates. For a verdict job, it is the decision endpoints plus the cheapest generator you already pay for, as a baseline.
Test the shortlist on your own cases
Launch benchmarks are chosen by the vendor, and Gemini Decision 1 shows why that matters. Google says on benchmarks spanning agent routing, reasoning, ranking, and multimodalmultimodalAI that works across several input and output types at once: text, images, audio, video and data, instead of one format only.View full definition → understanding, Gemini Decision 1 leads across the suite, with the lowest calibration error. The same day, the independent benchmarking firm Artificial Analysis put it ninth of 18 at launch. The overall rank also hides how uneven the model is. Gemini led the independent suite's knowledge and reasoning tests. On AA-Paperclip, which asks 15 questions about each résumé, research paper, invoice or receipt, Gemini got every answer right on 33% of documents, versus 35.5% for Jev and 63% for Drex Max v1.1. If your job is invoice checking, the knowledge-test lead tells you nothing about how it will do.
Fine-tuned models show the same pattern. Abacus.AI applied the same fine-tuningfine-tuningFine-tuning adapts a pre-trained model to a specific task or domain by continuing training on a smaller, targeted dataset, improving accuracy and style for that use case.View full definition → recipe to three very different open-weight models. The gains landed in completely different places, according to KDnuggets. Abacus.AI sells these models, so its own figures need independent checks. A model's average score does not predict how it will do on your particular job.
The fix is cheap:build a small test set of 20 to 50 real past cases where you already know the right answer, including the messy ones (scanned PDFs, angry emails, edge-case clauses). Run every shortlisted model on the same set. Score correctness, and for verdict models, check whether the high-confidence answers really are right more often.
Price the full task
Token prices mislead in both directions, so price the whole job, and includecost and latency in the same sheet.
A worked example, hypothetical and arithmetic only: a team screens 10,000 supplier documents a month, each about 3,000 tokens long. That is 30 million input tokens.
- On Gemini Decision 1 at the list rate of $0.10 per million input tokens, the documents cost $3 a month to read, before the tokens for the questions themselves.
- On Gemini 4 Argon, where pricing starting at $2 per million input tokens and $10 per million output tokens has been reported, input alone is $60.
- On Claude Fable 5.2, reported at $8 per million input tokens and $40 output tokens, input alone is $240, plus whatever the model writes.
The gap is roughly two orders of magnitude, but the cheap option only counts if the job really is a verdict and the accuracy holds on your test set. Per-task numbers from independent testing also differ from the list price. Across its index, Artificial Analysis estimates Gemini's cost at $345 per million tasks, versus $103 for Drex Max v1.1. Those are benchmark-specific task costs, not Google's advertised token rate or a universal deployment bill. Your own test set gives you your own number.
Keep a second vendor ready to swap in
Switching costs are falling at every layer, and Anthropic showed this at the infrastructure level this week. According to The Decoder, with Mythos, Anthropic brings new hardware to a usable performance level in about three days, instead of the six to twelve months traditional engineering approaches need to onboard new accelerators. Read the fine print: a footnote on the slide defines the three days as the approximate time for Claude Mythos Preview to bring up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance. The principle still applies to you. Anthropic treats compute from different vendors, chip architectures, and locations as one interchangeable pool. Do the same with models. Write prompts and test sets that do not depend on one vendor, so moving a job from one model to another means rerunning your test set rather than rebuilding the workflow.
Where model selection for a job goes wrong
Trusting the launch chart. Google's suite and the Artificial Analysis index tell different stories about the same model on the same day. Treat any vendor's benchmark table as a list of candidates to test.
Reading confidence as correctness. Confidence scores also aren't correctness warranties. On Artificial Analysis's calibration measure, where lower is better, it reports 0.18 overall for Gemini, compared with 0.16 for Drex Max v1.1 and 0.19 for Jev 1.14. It says calibration varies by task. Set your auto-accept threshold from your own test set, never from the launch blog.
Assuming the maximum specs hold. Gemini Decision 1 scores options without token-by-token decoding, keeping median latency 50-100 ms of model compute (150-250 ms end-to-end), across multimodal inputs and contexts up to 16K tokens, yet the API accepts far more. One analysis notes the figures are not a speed promise for every supported input size or all 1,000 questions. Time your own largest files.
Paying for paragraphs on a verdict job. Routing tickets, checking policies and grading drafts through a frontier chat model is the most common overspend. The vendors have spotted this too: OpenAI's pitch is that developers should stop paying for paragraphs when all they need is a verdict.
Giving a verdict model a reading job. The 33% all-correct rate on document questions is a warning for anyone hoping to extract invoice or contract fields with a decision endpoint. If the right answer depends on careful reading of a long document, test a generator alongside it.
Ignoring where the data goes. The same model can come through several routes with different terms. OpenRouter's listing for OpenAI's decision model, for example, says it is not currently available through EU in-region routing. Check where data is processed for the exact route you will use.
What to do on monday
- List the five AI tasks your team repeats most and mark each one verdict or prose using the whiteboard rule.
- For one verdict task, pull 30 past cases with known answers and turn them into a spreadsheet test set.
- Run that set through Gemini Decision 1, where developers can access the service with an API key in Google AI Studio, where a no-cost prototyping tier is available, and through OpenAI's Decisions endpoint, and log accuracy alongside confidence.
- For one prose task, send the same ten prompts to your current tier and one tier up, then ask the person who owns that output to rank them without seeing which model wrote what.
- Write the cost per completed task, not per token, next to each result, and set the confidence level below which a human reviews.
Within two weeks, all three vendors have launched new models, and none of those launches tells you which model fits your invoice queue or your client letters. Your 30-case test set does, and it keeps working: rerun it on each new launch in an afternoon instead of starting the debate again.
Frequently asked questions
Is ChatGPT, Claude or Gemini best for business work in 2026?
None of the three is best across all business work in 2026, because each vendor sells several tiers built for different jobs. Anthropic itself suggests Claude Opus 5.5 for daily tasks and Fable 5.2 for hard engineering problems. The reliable method is to run two or three candidates on 20 to 50 of your own past cases and compare accuracy and cost per task.
What is a decision model in AI?
A decision model scores a fixed set of answers you define in advance, such as yes/no, multiple choice or a rubric grade, and returns probabilities instead of writing text. Google's Gemini Decision 1 can score up to 1,000 such questions about one input in a single pass, and OpenAI offers a similar Decisions API built on GPT-6 Luna.
Can I trust the confidence score from Gemini Decision 1 or OpenAI's Decisions API?
Confidence scores from decision models are useful for setting review thresholds, but they do not guarantee correctness. Artificial Analysis reports that calibration varies by task, with Gemini Decision 1 scoring 0.18 on its Brier measure against 0.16 for Drex Max. Set the threshold for auto-accepting answers by testing on your own labelled cases.
How much cheaper is a decision API than a frontier chatbot model?
For closed-answer jobs, decision APIs can be roughly two orders of magnitude cheaper on input. Gemini Decision 1 lists $0.10 per million input tokens with no output charge, while Claude Fable 5.2 is reported at $8 per million input tokens. The saving only counts if the job has fixed answers and the decision model's accuracy holds on your own test set.
Go deeper
The lessons that take this article further, free to read.
- 1Cost, latency, and model selection tradeoffsBuilding with AI
- 2The Gemini model family: pro, flash, and when to use eachGemini & Google AI
- 3The OpenAI model family: GPT, reasoning models, and when to use eachChatGPT & the OpenAI ecosystem
- 4The Claude model family: opus, sonnet, haiku, and when to use eachClaude & the Anthropic ecosystem
- 5Evaluating outputs: how do you know it works?Building with AI
Sources
- Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass
- Gemini Decisions API Scores Yes/No, Choice and Rubric Questi
- OpenAI opens its Decisions API to all developers, pitching GPT-6 Luna as a fast, input-priced engine for yes-no, choice and scoring calls · AI Polix
- GitHub - jqueryscript/anthropic-claude-timeline: A public timeline of major Anthropic Claude model releases, product updates, and developer platform milestones. · GitHub
- Gemini 4 Argon: Release, Price and Access
- Google Drops Gemini Decision 1 to Score Choices 5x Cheaper Than Flash
- Anthropic releases first frontier models since Amodei's call for slowdown as IPO buzz builds
- Google AI news: October 2026 updates
- OpenAI announces a Decisions API on GPT-6 Luna, in limited preview — ModelSystem.One
- Gemini Decisions API: Fast, calibrated multimodal decisions
- Google’s Gemini Decision 1 model scores up to 1,000 questions in one API call
- Where Does Abacus.AI’s Smaug Model Improve Performance the Most? - KDnuggets
- Gemini 4 Nears Launch as Google Trails by 17 Points [2026]
- Anthropic releases new 'cheaper and faster' Claude models ahead of IPO
- Google’s Gemini Decisions API skips the chatbot essay · Digg
- Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days
- Anthropic Roadshow Video: The AMD and Disease Claims
- Google launches an AI decision API that scores up to 1,000 questions per request
- GPT-6 Luna Decisions - API Pricing & Providers
- Google Launches Fast, Calibrated Gemini Decisions API for Multimodal Decisions
- Google Deepmind's AlphaProof Code verified Signal's cryptographic core in four weeks instead of years
- Big Walk used AI to fix a bug. Now the indie gaming world is having a reckoning.
- Try Tabletop: Customizable, collaborative spaces for groups to plan, play and do more together.
- Plurai wants to help AI agents work better using the same approach that helped get self-driving cars to roads
- DeepL wants real-time translation to finally capture sarcasm and mood shifts
- 9 Key LLM Inference Metrics Explained in 5 Minutes
- UK regulator threatens universities with sanctions unless they adopt AI-proof exam formats
- Experience the art and legacy of Dr. Esther Mahlangu
- ChatGPT Plus for veterans and servicemembers
- OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect
Finished reading?
Validate your read to earn XP and feed your radar.