A New Leaderboard Takes Shape
The updated leaderboard, published July 8, features Claude Fable 5 at No. 1, followed by OpenAI’s GPT 5.5 at 80.2 percent and Claude Sonnet 5 at 76.2 percent. Google’s own Gemini 3.1 Pro Preview placed fifth at 73.7 percent, trailing GPT 5.4 as well. Other newly added models include Qwen 3.7 Max, Qwen 3.7 Plus, GLM 5.2, Kimi K2.7 Code, and MiniMax M3. developer.android.com
The results continue a pattern first noted in May, when GPT 5.5 topped the leaderboard and Gemini trailed OpenAI’s models by roughly two percentage points. With Fable 5 now in the mix, the gap at the top has widened further. The model, released by Anthropic in June, has posted leading scores across multiple coding benchmarks, including an 80.3 percent on SWE-Bench Pro. atlascloud.ai 9to5google.com thenewstack.io
Methodology Changes and Open Contributions
Beyond adding models, Google migrated the Android Bench framework to Harbor, which it described as enabling tougher, real-world AI reasoning tests. The benchmark evaluates models across 100 test cases run 10 times each, measuring not only accuracy but also latency and cost per run — dimensions Google first introduced in May. x.com android-developers.googleblog.com developer.android.com
Google also indicated that developers can now contribute to the benchmark directly, a shift that could broaden the range of Android-specific coding challenges used to test models. Several older models, including Claude Opus 4.6, Claude Sonnet 4.5, and earlier GPT variants, were archived in the update. developer.android.com x.com
What the Rankings Reveal
The cost and speed data in the new leaderboard offer a practical lens beyond raw accuracy. Deepseek V4 Flash, for instance, scores a modest 54.7 percent but costs just $1.50 per benchmark run, while Claude Fable 5’s top score comes at $133.20 per run. The tradeoffs underscore the choices facing Android developers as AI-assisted coding tools become embedded in everyday workflows — and as Google’s own models continue to chase rivals from Anthropic and OpenAI on the benchmark Google itself built. developer.android.com
