Google launches Gemini Decisions API with the gemini-decision-1 model, scoring up to 1,000 questions in one pass at $0.10 per 1M input tokens
- Gemini Decision 1 scores candidate options in a single forward pass instead of decoding token by token, returning probabilities and confidence scores at 150-250 ms end-to-end latency with no output tokens billed.
- One request takes a single multimodal input (text, images, audio, video, PDFs, up to 256K tokens) plus up to 1,000 independent questions of three types: yes/no, multiple choice, and rubric scoring.
- Pricing is $0.10 per 1M input tokens, which Google says makes it nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API across the shared benchmark suite.
- Google reports the lowest calibration error on benchmarks covering agent routing, reasoning, ranking and multimodal understanding, with median model compute of 50-100 ms that stays nearly flat from 1 to 100 questions per request.
- The endpoint ships today through the Gemini API in Google AI Studio with Python, JavaScript and REST guides; commenters ask when it reaches Vertex, since it appears to be AI Studio only for now.
Hacker News opinions
Looks like Google announced their own Jev-inspired API and model. Everyone seems to be converging on the same shape now.
Bit more expensive than the other decision models in the same lane, but it takes audio and video input too, and for those cases it's extremely useful.
Image input is the reason I never moved off my home-rolled Jev clone. Resize to 1k, ask a handful of tagging questions, roughly 3 images a second on Qwen 3 VL.
Meanwhile they're suffocating Jules today with no migration path beyond a one-line banner saying use Antigravity. Hard to believe this one survives six months.
That doesn't track. This is a model release competing with Jev, like Carbon and Argon. Models don't get killed without something to succeed them.
Barring Search, Gmail and YouTube I'm really wary of building on anything Google ships now.
Antigravity is far superior to Jules anyway. Plant a thousand flowers and cull the ones that don't bloom.
OpenAI took two weeks after Jev to ship a decision endpoint, Google took two months. Bureaucracy or quality control?
Their benchmarks actually put it ahead of OpenAI and Jev, and Jev somehow looks worse than itself in those same evals.
I'm surprised it took Jev for everyone to realize this is useful. People tried grammar-constrained JSON output years ago but you can't get probabilities out of the model that way. So how do these models compute them?
You can pull it from the logprobs of yes/Yes/YES versus no, by hand or with a library on top of vLLM. In Jev space I think the probability type is called noul.
Short answer is cross entropy against ground truth probabilities, same as any other classifier. The long answer is more and better data.
We were doing zero and few-shot classification with LLMs before they could reliably finish a sentence. The model just needed tuning for classification instead of generation.
So is it over for Jev now that the big boys are in the game?
They raised close to a billion after the OpenAI Decisions API shipped, so somebody thinks they have a next move.
Last time I checked Jev was still significantly faster. 250ms versus 750ms opens up a lot of use cases.
Anyone know when this lands in Vertex? As far as I can tell it's AI Studio only for now.
Best decision model I've used besides Jev is span-02, finetuned from Gemma4 26A4. Makes me optimistic despite the price.
This is Google buying share of wallet from its existing customers with a solid but not great me-too product. Standard practice from the big G.