Microsoft and OpenAI shipped their score-it-don't-write-it models within the last five weeks, and Google just answered with one that reads video and audio too. If you pay for classification or routing calls, the price you're quoted is about to matter less than what the model can look at.
Google launched the Gemini Decisions API on Wednesday, November 11, powered by a new model called Gemini Decision 1 and available in Google AI Studio. Sundar Pichai announced it on X, describing a model built for closed-set choices that scores up to 1,000 questions across text, images, audio, video and PDFs in a single pass, instead of generating text token by token. For the so-called system one models, the ones that were supposed to own this job, it's a demolition.
Frankly, that is the reading here, and it's an argument rather than a benchmark result. Google hasn't published a head-to-head against its rivals, and RuntimeWire noted the launch coverage doesn't include benchmark comparisons with OpenAI or Microsoft. What follows is the case from the facts that are public.
Parallel.ai defines a System One model as one that reads a state, such as a support ticket, a record or a web page, and returns calibrated probabilities over answers you define in advance. The name comes from Daniel Kahneman's fast-thinking System 1. It doesn't write anything, which is the point: agent systems stop paying for a paragraph when all they need is a branch, a rank or a score.
The category is barely a month old. OpenAI's Decisions API, built on GPT-6 Luna, went into public beta on October 6 at $0.10 per million input tokens, according to the comparison site systemonemodels.org. Microsoft released Decision-1 on October 9 through Foundry at $0.042 per million input tokens. Smaller names such as Jev were already in the mix.
Google Delays Gemini 3.5 Pro Launch to July 17 After Scrapping Its Base Model Google scrapped the original Gemini 3.5 Pro base model and restarted pretraining, pushing the launch to July 17, 2026. Leaked specs point to a 2 million token context window and a new Deep Think reasoning mode, landing right as DeepSeek's own July 24 deadline forces developers to migrate off its old model names. - when is Gemini 3.5 Pro launching - why Google delayed Gemini 3.5 Pro
Neither of the big two charges for output tokens, and neither had a story about audio or video.
Why Gemini Decision 1 reads like a wipe-out #
Start with the input. Gemini Decision 1 takes text, images, audio, video and PDFs. Systemonemodels.org says it's the only one of the four hosted models whose documentation lists audio and video. A support team that wants to flag angry calls, or a trust-and-safety team that wants to score uploaded clips, no longer has to transcribe first and then call a text-only scorer.
Then the volume. One request can carry up to 1,000 questions, and Interesting Engineering reported that the API returns probability and confidence scores in roughly 150 to 250 milliseconds, with decision types covering yes/no, multiple choice up to 255 options, and rubric scoring. It also reported a 256,000-token context. Run one document past a thousand checks and you pay for reading it once.
Price is the weak spot in the argument, and it should be said plainly. Gemini Decision 1 costs $0.10 per million input tokens with no output charge, the same as OpenAI and well above Microsoft's $0.042. Google is also offering a free tier in AI Studio, and Mezha reports integrations rolling out with OpenRouter, LangChain, Vercel, LiteLLM and Pydantic AI. Matching OpenAI's price while adding two media types and the largest question batch isn't a price war. It's a feature war, and the rivals are behind on features.
RuntimeWire's read of the launch is that Google wants the decision layer that sits in front of larger models, a high-traffic spot that compounds across every app call. That fits. Whoever scores the request first decides which bigger model, if any, gets called afterward.
What it doesn't prove yet #
Calibration is the whole product here, and nobody outside Google has tested whether a confidence of 0.9 means right nine times in ten on your data. Microsoft's cheaper model may still win for teams that only ever score short text, where $0.042 against $0.10 is a 2.4x gap that adds up at millions of calls a day. If you're picking one this week, run your own labelled sample through all three before you move anything into production.
Google out-featured the category within five weeks of its October launch.
Google buys the Wyoming data center Crusoe couldn't finish on its own Google will pour roughly $17 billion into the Cheyenne campus it just took over, after quietly pushing Crusoe out as developer since June over cost and schedule concerns, Bloomberg reported. - google acquires crusoe wyoming data center project - why google bought crusoe's cheyenne data center site
Also read: United Airlines bans AI agents from booking flights ahead of the holidays • Meta infrastructure chief Santosh Janardhan will leave in early 2027 as Dave Brown takes over • Fulcra CEO Michael Tiffany on Rent the Reasoning, Own the Context, and Why AI Agents Should Talk Across Platforms
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.