Google has officially updated its Android Bench AI programming benchmark, introducing results for three new models: OpenAI GPT-6.1 Terra, Kimi K3.1 Pro, and DeepSeek V4.1 Flash. This benchmark evaluates how effectively AI models handle real-world engineering tasks, including application creation, feature development, and code migration.
The Kimi K3.1 Pro model, utilizing the kimi-code agent, achieved a 24.0% pass rate across 30 long-range tasks. It recorded an average completion rate of 82.3% with an average execution time of 27.4 hours. The cost for a single benchmark run for this model is $317.4.
In terms of performance, Kimi K3.1 Pro outperformed GPT 6 Astra, Claude Fable 5.1, and Gemini 4 Argon, though it trailed behind GPT-6.1 Terra, which reached a 28.7% pass rate.
DeepSeek V4.1 Flash 1017, using the opencode agent, demonstrated a 17.3% pass rate and a 71.2% average completion rate. The model completed tasks in an average of 12.0 hours at a significantly lower cost of $9.9 per run, positioning it as a viable option for teams prioritizing speed and budget efficiency.
