Gemini

Gemini is the family of multimodal models built by Google DeepMind, Google’s AI lab. Google introduced it on December 6, 2023 as its “most capable and general model yet”, with a first version, Gemini 1.0, in three sizes: Ultra, Pro and Nano, the last for on-device tasks. Google said it designed Gemini to be “natively multimodal”, pre-trained from the start on different modalities, and developers could reach Gemini Pro through the Gemini API from December 13, 2023.¹ As of October 13, 2026, the Gemini API lists Gemini 3.8 Flash as its newest stable model, alongside Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.1 Flash-Lite, with Gemini 3.1 Pro in preview.² Google calls Gemini 3.8 Flash its most intelligent Flash model, built for long-horizon software engineering and autonomous agents; it takes up to 1,048,576 input tokens and returns up to 65,536, and its page was last updated in September 2026.³

On ARC-AGI-3’s Semi-Private set, run by ARC Prize on its Standard harness, Gemini 3.8 Flash scores 10.37% at High effort, fourth among 19 models and the highest from outside OpenAI and Anthropic; on the older ARC-AGI-2 it scores 89.2% (Highest ARC-AGI-3 Scores, by Model).⁴⁵ On SWE-bench Verified, Gemini 3 Flash resolved 75.8% of tasks in the shared mini-SWE-agent v2.0.0 harness, tied for second with MiniMax M2.5, at an average cost of $0.36 per task, and Gemini 3 Pro 69.6% (SWE-bench Verified, Pro and What Replaced Them).⁶ On price, Gemini 3.1 Flash-Lite lists $1.50 per 1M output tokens, thinking tokens included, and Gemini 3.1 Pro Preview $12.00 for prompts up to 200,000 tokens (Cheapest LLM APIs per 1M Output Tokens).⁷