ModelsArchitectures & capability
Google launches Gemini Decisions API, a multimodal model that returns probability-scored answers instead of text
Google's Gemini Decision 1 returns typed answers with probabilities instead of generated text. It accepts text, images, audio, video and PDFs, and charges only for input. TypeSafe's Jev and OpenAI already offer this kind of model, and research on Jev shows its cheap verdicts share LLM judges' blind spots.
From the Simon Willison article of November 11, 2026 and cited sources: Gemini Decision 1 (gemini-decision-1) answers yes/no questions, multiple-choice questions with 2 to 255 options, and ordered rubric scoring. [2] [8] [1]
SOURCES[2] Google: Gemini Decisions API blog.google[8] SiliconANGLE: Google enters decision model market siliconangle.com[1] Simon Willison: Gemini Decisions API simonwillison.net
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Gemini Decision 1 (gemini-decision-1) answers yes/no questions, multiple-choice questions with 2 to 255 options, and ordered rubric scoring. [2] [8] [1]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 11, 2026 article published
- Covered
- Nov 12 edition
- Evidence
- 9 sources across 8 sites
- 1Simon Willison: Gemini Decisions APIsimonwillison.net
- 2Google: Gemini Decisions APIblog.google
- 3NYU Shanghai RITS: TypeSafe AI Launches Jevrits.shanghai.nyu.edu
- 4Sun et al., arXiv 2609.26758arxiv.org
- 5Rao & Callison-Burch, arXiv 2609.29769arxiv.org
- 6Unite.AI: OpenAI Releases Decisions API in Public Betaunite.ai
Reach: not measured yet
Gemini Decision 1 (gemini-decision-1) answers yes/no questions, multiple-choice questions with 2 to 255 options, and ordered rubric scoring. It scores the candidate options directly instead of decoding tokens. Google says it can answer up to 1,000 questions over one shared input in 150–250 ms. Pricing is $0.10 per million input tokens with free output. In one informal hands-on test, it returned graded probabilities, while Jev and OpenAI's endpoint returned only 1.0 or 0.0. [2] [8] [1]
Google is the third vendor in this category, after TypeSafe's Jev and OpenAI's GPT-6 Luna decision endpoint. The three use nearly identical question schemas, so porting code between them should be simple. Gemini differs on input types: it is the only one of the three that takes audio, video and PDFs directly. Jev handles text and JSON only, and OpenAI's endpoint takes text and inline images. [1] [6] [3]
Two arXiv studies examined Jev. Decision models cost far less than LLM judges, which cost 16 to 357 times as much. But their mistakes largely match the mistakes LLM judges make, so falling back to an LLM recovered at most 2.7 points. Renaming answer options with yes/no labels dropped mean AUC from 93.8% to 23.2%, while every output stayed well-formed. [5] [4]
free output and input prices of 4 to 10 cents per million tokens make high-volume routing, moderation and grading much cheaper than using LLM judges.
Read the full assessment
Implication: teams can move classification off general models, but Google's benchmark leads are self-reported and should be checked against in-house labels.
Executive brief
Google is the third company to ship a "decision model": an API that returns typed answers with probabilities, not generated text. The interface of Gemini Decision 1 (gemini-decision-1) is close to that of TypeSafe's Jev. Google charges $0.10 per million input tokens and nothing for output, the same rate as OpenAI's GPT-6 Luna decision endpoint (Simon Willison). Where it differs is input types. Google says it accepts text, images, audio, video and PDFs, and can answer up to 1,000 questions in one pass in 150–250 ms (Google). Independent research on Jev suggests these models are cheap but share LLM judges' blind spots.
What changed and event timeline
TypeSafe launches Jev
Jev is a "System One" model that returns decisions instead of free text. TypeSafe was founded by former OpenAI researcher Diogo Almeida, and the launch came after two years in stealth ().
OpenAI Decisions API enters beta
It runs on gpt-6-luna through
POST /v1/decisions, accepts text and images, and OpenAI says it is about 10x faster than the Responses API ().OpenAI opens the endpoint to all developers
This was reported alongside the beta coverage, with general availability expected within weeks ().
Google launches the Gemini Decisions API
It is available in AI Studio with a free prototyping tier. Google lists OpenRouter, LangChain and Vercel among its integrations (;).
First hands-on test
Willison asked yes/no questions about a video of a robot. Gemini returned probabilities around 0.92 for the true answers and 0.05–0.08 for the false ones ().
Capabilities and access
- Model: Gemini Decision 1 (
gemini-decision-1), through the Gemini API in Google AI Studio (Google). - Question types: yes/no with optional criteria, multiple choice with 2–255 options, and ordered rubric scoring.
- Inputs: text, images, audio, video and PDFs. The post does not reconcile the two.
Read the full section
- Model: Gemini Decision 1 (
gemini-decision-1), through the Gemini API in Google AI Studio (Google). - Question types: yes/no with optional criteria, multiple choice with 2–255 options, and ordered rubric scoring.
- Inputs: text, images, audio, video and PDFs. Google's post cites inputs up to 256K tokens and also a 16K-token figure for multimodal inputs; the post does not reconcile the two.
- Pricing: $0.10 per million input tokens, free output, and no separate charge for transcription or OCR.
- Competitors: Willison reports Jev at 4.2 cents per million tokens and text/JSON only. OpenAI's endpoint takes text and images, with images sent as inline base64 only (Unite.AI).
Technical analysis for researchers and developers
- How it answers: Google DeepMind says the model scores the candidate options directly instead of decoding tokens, with median model compute of 50–100 ms (SiliconANGLE).
- Not disclosed: architecture details, training method and benchmark protocols.
- Interface convergence: the three vendors use near-identical question schemas. Porting code between vendors should be straightforward.
Read the full section
- How it answers: Google DeepMind says the model scores the candidate options directly instead of decoding tokens, with median model compute of 50–100 ms (SiliconANGLE). Google says latency stays nearly flat from 1 to 100 questions per request, because all questions share one encoded input (Google).
- Not disclosed: architecture details, training method and benchmark protocols.
- Interface convergence: the three vendors use near-identical question schemas. Gemini even reuses Jev's term "Noul" for a yes/no question (Simon Willison; Sanity glossary). Porting code between vendors should be straightforward.
- Output handling: treat outputs as calibrated scores and set your own thresholds rather than reading them as booleans.
Claims and evidence
- Google says Decision 1 leads on routing, reasoning, ranking and multimodal benchmarks and has the lowest calibration error (Google). These results are vendor-reported.
- "Nearly 5x cheaper" than Gemini 3.5 Flash-Lite: this is also Google's own figure.
- Graded rather than binary answers: Willison saw Gemini return graded probabilities, while Jev and OpenAI returned 1.0 or 0.0 in his tests (Simon Willison).
Read the full section
- Google says Decision 1 leads on routing, reasoning, ranking and multimodal benchmarks and has the lowest calibration error (Google). SiliconANGLE reports a 7-of-9 benchmark win rate (SiliconANGLE). These results are vendor-reported. No independent replication was found.
- "Nearly 5x cheaper" than Gemini 3.5 Flash-Lite: this is also Google's own figure.
- Graded rather than binary answers: Willison saw Gemini return graded probabilities, while Jev and OpenAI returned 1.0 or 0.0 in his tests (Simon Willison). This is a single informal test. OpenAI's documentation describes its predicate output as a probability between 0 and 1 (Unite.AI).
Context and prior work
- Jev's traction: Vercel reported that Jev reached more than twice as many paid teams in 24 hours as any earlier model launch on its AI Gateway (NYU Shanghai RITS).
- Funding: SiliconANGLE reports TypeSafe raised $870M at a $7.5B valuation in October (SiliconANGLE).
- Precedent: the idea of an LLM-as-judge returning constrained, classifier-style outputs is older. Perplexity's Decider was already being benchmarked against Jev in September (arXiv 2609.29769).
Read the full section
- Jev's traction: Vercel reported that Jev reached more than twice as many paid teams in 24 hours as any earlier model launch on its AI Gateway (NYU Shanghai RITS).
- Funding: SiliconANGLE reports TypeSafe raised $870M at a $7.5B valuation in October (SiliconANGLE).
- Precedent: the idea of an LLM-as-judge returning constrained, classifier-style outputs is older. Perplexity's Decider was already being benchmarked against Jev in September (arXiv 2609.29769).
Limitations, safety and contested findings
- Shared blind spots (Rao and Callison-Burch): decision models cost far less than LLM judges, which cost 16–357x as much.
- Option names override definitions: relabeling options with yes/no names cut mean AUC from 93.8% to 23.2%, which is below chance.
- Scope of evidence: both studies tested Jev, not Gemini Decision 1.
Read the full section
- Shared blind spots (Rao and Callison-Burch): decision models cost far less than LLM judges, which cost 16–357x as much. Their errors largely match LLM judges' errors: over 85% of LLM verdicts on decision-model mistakes repeated those mistakes, and cascading to an LLM fallback added at most 2.7 points (arXiv 2609.29769).
- Option names override definitions: relabeling options with yes/no names cut mean AUC from 93.8% to 23.2%, which is below chance. Type-error rates stayed at 0% throughout. A well-formed answer is not necessarily a correct one (arXiv 2609.26758).
- Scope of evidence: both studies tested Jev, not Gemini Decision 1.
- OpenAI constraints: OpenAI's endpoint cannot chain dependent decisions within a single request (Unite.AI).
Business and practitioner implications
- Where it fits: Google's own uses include model routing in Gemini Notebook (which Google says cut latency on routine queries by 40%) and a safety layer for MCP tools (Google).
- Price pressure: with free output and input prices of 4–10 cents per million tokens, high-volume classification moves away from general LLMs.
- Multimodal gap: Gemini is the only one of the three that takes audio, video and PDF directly, which matters for media moderation.
Read the full section
- Where it fits: Google's own uses include model routing in Gemini Notebook (which Google says cut latency on routine queries by 40%) and a safety layer for MCP tools (Google). The same pattern suits routing, policy checks, content moderation and grading.
- Price pressure: with free output and input prices of 4–10 cents per million tokens, high-volume classification moves away from general LLMs.
- Multimodal gap: Gemini is the only one of the three that takes audio, video and PDF directly, which matters for media moderation.
- Validation: check the models against your own human labels, and use neutral option names such as 0/1 to avoid the naming bias described above.