SIAgentNews.com
← All news

Google's Android coding benchmark now lists Kimi K3.1 Pro and DeepSeek V4.1 Flash

Google's Android Bench now has scores for Kimi K3.1 Pro and DeepSeek V4.1 Flash, showing a clear trade-off between accuracy, speed and cost.

· First reported by ifeng Tech

24.0%Kimi K3.1 Pro pass rate on Android Bench

The short version

  • Google's @AndroidDev account posted the first Android Bench scores for three models on November 10.
  • Kimi K3.1 Pro: 24.0% pass rate, 82.3% average completion, 27.4 hours, $317.4 per run.
  • DeepSeek V4.1 Flash: 17.3% pass rate, 71.2% average completion, 12.0 hours, $9.9 per run.
  • GPT 6.1 Terra leads Kimi with a 28.7% pass rate; Kimi beats GPT 6 Astra, Claude Fable 5.1 and Gemini 4 Argon.

Why it matters

If you are choosing an AI model for Android work, this ranking measures real app-building tasks rather than general chat skill. It also shows that the cheapest option can be much faster and cheaper while passing fewer tasks. The right pick depends on your budget and deadlines.

What Google published

As ifeng Tech reported, Google's Android Developers account posted on X on November 10 with the first public Android Bench scores for three models: OpenAI's GPT-6.1 Terra, Kimi K3.1 Pro and DeepSeek V4.1 Flash.

Android Bench is Google's coding benchmark for Android work. It tests whether an AI can handle real engineering jobs: creating apps, building features, migrating code and converting apps.

Kimi: higher score, higher bill

Kimi K3.1 Pro ran with an agent called kimi-code, meaning software that lets the model plan and take steps on its own. It was tested on 30 long-running tasks and passed 24.0% of them, with an average completion of 82.3%.

Each task averaged 27.4 hours, and one full benchmark run cost $317.4. That score beats GPT 6 Astra, Claude Fable 5.1 and Gemini 4 Argon, but trails GPT 6.1 Terra at 28.7%.

DeepSeek: the budget option

DeepSeek V4.1 Flash, listed in the source as "V4.1 Flash 1017", was run through an agent called opencode. It passed 17.3% of tasks, with 71.2% average completion.

It averaged 12.0 hours per task, and a full run cost $9.9. The source says its speed and cost make it a good fit for teams watching budget and turnaround time.

How to read the numbers

Even the top score here, 28.7%, is below one in three tasks passed. These are hard, long jobs, and the models are not close to finishing everything.

Each model also ran with a different agent, so the comparison is not purely model against model. Treat the ranking as a starting point and test on your own project.

What to watch

  • Will Google add more models or publish the full task list and scoring details?
  • Do pass rates rise when models use the same agent setup?

Try it

Find the Android Developers post on X and check the task list and model versions. Then run one real task from your own app through your top two candidates.

Sources and links

First reported by ifeng Tech. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news