Lilith.
⌕
Editorial illustration: Google puts Gemini Decision 1 on fast choices where a large model is overkill
Lilith illustration · editorial remix

Google has entered the decision model category with Gemini Decision 1, built for yes or no answers, multiple choice and rubric scoring. According to the Latent Space roundup, it accepts text, images, audio, video and PDFs, handles up to 1,000 questions in one request and returns a confidence score for every answer.

Decision 1 scores choices in one pass instead of generating a long answer

The model works with a closed answer set and can score up to 255 options. Its stated context window is 256K and end-to-end latency ranges from 150 to 250 ms. Pricing is $0.10 per million input tokens with no charge for output.

That changes the economics of tasks where an application needs a dependable choice rather than a paragraph: route a ticket, select a tool, score a document or decide whether to escalate to a person. Latent Space also reports LangChain support and a Stitch experiment that turns a rough prompt into a multi-screen product flow in about 5 seconds.

The largest saving comes from pairing a fast switch operator with an expensive diagnostician

At AI Engineer, Cloudflare demonstrated a hybrid pattern in which a decision model selects tools and extracts relevant logs while a large reasoning model handles diagnosis. In the reported scenario, LLM calls fell from 6 to 1 and token use from about 40K to 1.8K.

For product teams, this is a more useful frame than replacing one chat model with another. Most agent workflows contain many small bounded decisions and a few problems that genuinely need generative reasoning. A specialist can remove cost and latency from the first group.

Cheap tokens conceal a less flattering cost per completed task

Google says the input price is about 5 times lower than Flash-Lite, but the comparison cited by Latent Space is less comfortable. Artificial Analysis scored Decision 1 at 66 on its Decision Index, tied with Jev 1.14 and Elephant 9B, while Drex Max v1.1 and GLiDE v1.2 were 7 points higher.

Its calculation put Decision 1 at $345 per million tasks, compared with $103 for Drex Max and $156 for Jev. Average latency was 211 ms, also slower than Jev at 92 ms. These are results from a particular benchmark rather than a universal ranking, but they show why token pricing alone is inadequate.

A team’s own error matrix will matter before the next public leaderboard

Teams should measure false approvals, needless rejections and confidence calibration on their own data. In ticket routing, a cheap mistake may move one message to the wrong queue. In payment blocking or a security intervention, the same mistake can cost far more than thousands of inferences.

Behaviour under data shifts and the ability to express uncertainty will matter as well. A decision model belongs on the critical path only after the team knows the price of its errors and sets a threshold beyond which a person or stronger model takes over.

Lilith's verdict

A decision model belongs at the switch, not in the driver’s cab. If it cheaply sends the wrong case down the wrong track, a $0.10 per million token price will comfort nobody.