Gemini Decision 1, OpenAI Decisions API, Microsoft Decision-1: Among the two devices with the same price, Google does not cache

Google's access documentation contains a trap that's easy to step on. You classify the work order, write the options as shipping, billing, none_of_the_above, and plan to let the third person take the lead. As a result, the option keys are sorted alphabetically before scoring the model, and none_of_the_above is ranked before shipping.

What’s even more troublesome is the second one: choice questions will always assign probabilities to the options you listed, even if none of them are suitable. A message asking "What time does your store open on weekends?" will still be classified as "most like billing", and the probability may not be low.

The document's own suggestion is to rename the hidden item other, and add a separate yes or no question: "Is this message related to the order?"

This is Gemini Decision 1, which went live on November 11th.

It is the same type of thing as the Microsoft Decision-1 that came out a month ago and the OpenAI Decisions API that entered public beta in early October: no text is written, only probabilities are assigned to options, and it is used for judgment nodes in the agent processes such as routing, review, and scoring. The October article on the site has already compared Jev, Clef and Microsoft. This time Google enters the market and the accounts have to be recalculated.

Put three price lists together

The unit prices of Google and OpenAI are exactly the same, both about 2.4 times that of Microsoft.

If your input is all plain text work orders, there will be no reward for this difference. Google itself has done a comparison, leading Jev and OpenAI in seven out of nine tests, but the same comparison also admitted that in terms of cost efficiency alone, Jev is slightly better. Someone in the Hacker News post said it very directly: It is relatively expensive in the same field. The advantage is that it can consume audio and video, and it will be very useful in those types of scenarios.

It’s actually very clear where to spend more money: multi-modality and large windows.

There is no cache. This is more important than the unit price.

The billing rule of Decision 1 has a favorable design: in the same request, the shared input is only counted once, and for each additional question, only the description text of the question itself is added. Asking the same work order ten questions is much cheaper than making ten adjustments.

The downside is hidden in the "not supported" list. Context caching, Batch API, Flex stalls, none of them now.

Be specific. Assume that your audit rule document contains 20,000 tokens. You must bring it with you every time. If you run 10,000 requests a month, the rules alone are 200 million tokens. At $0.10 that’s $20. The same fixed prefix is ​​cached on Gemini 3.5 Flash-Lite, and the cache hit price is $0.03, which is only $6, a difference of more than three times. Of course, Flash-Lite has to pay an output fee. Someone calculated it using 10,000 work orders with 600 tokens each and asking 3 questions each: Decision 1 costs about US$0.78 a month, and Flash-Lite plus structured output costs about US$3.84, not counting the thinking tokens. So the conclusion is divided into two parts: short prefix, many problems, Decision 1 is cheaper; long prefix, fewer problems, the disadvantage of no caching will eat up the unit price advantage.

Google's claim of "nearly 5x cheaper than Flash-Lite" is an average on its own benchmark suite.

Latency numbers only cover up to 16K

The latency given by Google is 150 to 250 milliseconds end-to-end, of which the model takes 50 to 100 milliseconds, which is almost unchanged from 1 question to 100 questions.

Note the premise of this set of numbers: the context does not exceed 16K tokens.

The window was opened to 256K, but the speed test only measured the first 6%. If you use it to judge a contract PDF of several hundred pages, or a long recording, the official has not said what the delay will be. The selling point is a large window, but the promised speed is only valid in a small window. This is mentioned briefly in the release.

Microsoft reported an accuracy of 83.5% on nearly 150,000 questions in 85 milliseconds and 36 benchmarks. The calibers of the two are completely different. One is end-to-end, and the other does not specify the measurement range, so the speed cannot be directly compared.

Microsoft is cheaper, but the interface and base are variables

The price of Decision-1 is the lowest among the three, $0.042, which is the same price as Jev. Its shortcomings were written last time, and they still hold true this time: it does not use the OpenAI-compatible Chat Completions interface, there is only one Microsoft hosting on OpenRouter, and there is no backup provider. The base is trained with Qwen3.5-9B. Microsoft said that it will switch to MAI and OpenAI models for retraining in the future. The thresholds adjusted today may need to be readjusted by then.

It gave a beautiful robustness figure: only 1.3% of judgments were flipped on average under eight types of perturbations, and the options were not flipped once when the wording was changed or the order was disrupted. According to the description of "sort options alphabetically first" in Google's documentation, the order problem has been solved by both companies, but Google's way of handling it will change the position of your pocket options.

The 32K window, coupled with the lack of explicit multi-modal support, makes it suitable only for short text.

OpenAI has the least information

OpenAI's Decisions API is built on GPT-6 Luna, has the same price as Google, and can read text and embed images. After entering more than 272,000 tokens, the price doubles, which is in direct contrast to Google's large window. It is the one being compared in Google's nine-point comparison, and independent testing has not yet seen it.

The community’s ridicule of the timeline is quite accurate: OpenAI followed Jev two weeks after it came out, and Google was two months behind. The latecomer can only rely on multi-modality to round out the price difference.

Google’s own shortcomings

In addition to no caching and only measuring 16K latency, there are three other things:

First, the custody channels are narrow. No weighting, no self-deployment, "coming soon" on Vertex AI but no date. Currently you can only use Gemini developer API and OpenRouter.

Second, the data in the free tier will be used by Google to improve products, but the paid tier will not. When using real customer data to prototype on the free tier of Google AI Studio, you must first think clearly about compliance.

Third, the document specifically warns: User input may "defend their own classification." A refund request that says "This is obviously a quality issue" may have convinced the model. When it comes to judgments about compensation and refunds, Google recommends leaving manual work or a second check.

It does something that others can't do. In Google's example, a headphone work order comes with order information, and three questions are asked together: Is it a hardware failure? 0.94, what should I do? Manual upgrade is selected, 0.76, and the risk of churn is 2.99 points (out of 3). The entire request requires 324 input tokens. Replace it with a customer service recording plus account information, a scanned claim form plus photos of the damage, and it's the same call without having to pay for transcription and OCR.

Gemini Notebook uses it internally to assign model grades and thinking depth to each request. Google said that daily query latency has been reduced by up to 40%.

Select by what you input looks like

The input is plain text, large in volume, and short in context: Microsoft Decision-1 or Jev, the price is as low as more than 40% of the original price.

The input includes audio recordings, photos, and PDFs: currently only Decision 1 is the only hosting option that can natively consume it.

The fixed rules document is very long and needs to be carried every time: calculate the cache account first, don’t just look at the unit price.

It is already in the OpenAI system and only requires text and screenshots: its Decisions API has the lowest migration cost, and it will be more stable to go into production after the public beta is over.

Almost all the results of the three companies come from the manufacturer's self-test. Before going online, run it through two to three hundred real samples with manual annotation. It is more reliable than any comparison chart. The tools are already available: LiteLLM is connected to Decision 1 through the /v1/decisions interface, and it can also be adjusted directly on OpenRouter.

Related comparison: "Decision-1, Jev, Clef: Three models that only do multiple-choice questions, Microsoft has set the price exactly the same as Jev" | "Muse Ember, Claude Haiku 5.5, GPT-6 Luna: The price is the same, but the difference in the account for each question is almost three times"