Google's own Android benchmark: Gemini costs 17.5 times as much as DeepSeek V4 Flash
Google Android Bench list shows that the single running cost of Gemini 3.5 Flash is $147, while DeepSeek V4 Flash is only $8.4, a price difference of 17.5 times. GPT-5.5 ranked first with a score of 74, and domestic models GLM, Kimi, and DeepSeek collectively made the list.
Google’s own benchmark test overturned: Gemini cost 17.5 times that of DeepSeek
Google has updated the Android Bench list - a benchmark test that specifically measures the ability of AI models to complete Android development tasks, which has a strong reference value for developers to select models. The results are quite ironic: Gemini 3.5 Flash, which Google high-profile claimed at I/O 2026 as "the most powerful Flash model to date," only scored 63.7 points and ranked sixth. The average cost of a single run is as high as $147.1 (approximately 996 yuan), making it the most expensive model in the entire list. The single-time running cost of DeepSeek V4 Flash is only $8.4 (approximately 57 yuan) - the former is 17.5 times of the latter.
List results and cost details:
| Ranking | Model | Score | Cost per trip |
|---|---|---|---|
| 1 | GPT-5.5 | 74.0 | — |
| 2 | GPT-5.4 | 72.4 | — |
| 3 | Gemini 3.1 Pro Preview | 72.4 | ~$49 |
| 4 | Claude Opus 4.7 | 68.7 | — |
| 5 | Claude Opus 4.6 | 66.6 | — |
| 6 | Gemini 3.5 Flash | 63.7 | $147.1 |
| 7 | GLM 5.1 | 59.7 | — |
| 8 | Kimi K2.6 | 58.6 | — |
| 10 | DeepSeek V4 Pro | 55.4 | — |
| 12 | DeepSeek V4 Flash | 52.7 | $8.4 |
In terms of domestic models, GLM 5.1 (59.7 points), Kimi K2.6 (58.6 points), and DeepSeek V4 Pro (55.4 points) were on the list at the same time, reflecting the collective presence of Chinese AI models in the field of developer tools.
Google had claimed at I/O 2026 that Gemini 3.5 Flash encoding capabilities were more robust, with output speeds up to 4 times that of competing leading-edge models, and outperformed Gemini 3.1 Pro in some internal benchmarks. However, Android Bench's real Android development scenario data gives a completely different conclusion - not only does the score not make it into the top five, but the price/performance ratio is even crushed by DeepSeek V4 Flash. Even Google’s own benchmarks can’t save face for its own models. In an era where developers vote with their feet, manufacturers’ marketing rhetoric is being systematically falsified by cost-effective data from real scenarios.
Reviews