Google: Gemini 3.5 Flash

Gemini 3.5 Flash is well supported by official Google and DeepMind sources: model code, 1,048,576-token input limit, 65,536-token output limit, native multimodal inputs, and Gemini API features including function calling, structured outputs, code execution, grounding, caching, Batch API, and thinking. This justifies high technical, multimodal, and developer-experience scores, especially for API products that need long-context multimodal I/O rather than only chat quality. Public coding evidence is positive but not frontier-leading. DeepSWE/DataCurve lists gemini-3.5-flash[medium] at about 28% Pass@1, with reported average cost and time, and evals.report repeats a 28.32% official DeepSWE result. LiveCodeBench evidence is mixed: the official leaderboard excerpt did not visibly list Gemini 3.5 Flash, while LayerLens reports a high Stratix LiveCodeBench result from a third-party evaluation. Google’s own evaluation note references Terminal-Bench 2.1 methodology, but the supplied excerpt does not provide a comparable score here. Pricing is clear from Google and Artificial Analysis at $1.50/M input and $9.00/M output, with batch/flex/priority tiers noted. Artificial Analysis reports strong throughput but high TTFT depending on measurement, so speed is good but not uniformly low-latency. LMArena hard-prompts rank 22 with 1492±8 and 6,785 votes is useful public preference signal, not a substitute for task-specific evals. Overall evidence quality is good, with remaining caution around third-party benchmark comparability and incomplete official leaderboard coverage.