Google Launches Gemini 3.1 Pro
Google has launched Gemini 3.1 Pro, a new model with a strong focus on software engineering, coding, and autonomous agents. The release positions Gemini as a direct competitor to specialized coding models like Codex and Opus, underscoring AI-driven development as a top priority for Google.
Released on February 19, 2026, Gemini 3.1 Pro represents a significant upgrade in reasoning, scoring 77.1% on the ARC-AGI-2 benchmark, which tests the ability to solve novel logic patterns. This is more than double the reasoning performance of its predecessor, Gemini 3 Pro. The model's architecture now supports a 1 million token context window and, crucially, an expanded output capacity of 65,536 tokens. This directly addresses a limitation in the previous version, allowing for the complete refactoring of large code files in a single pass without truncation. In the competitive landscape, Gemini 3.1 Pro establishes a new record for competitive programming with an Elo of 2887 on LiveCodeBench Pro. However, Anthropic's Claude Opus 4.6 still leads on the SWE-Bench Verified benchmark for resolving real-world software issues, while OpenAI's GPT-5.3-Codex excels at terminal-heavy workflows. A key differentiator is cost-efficiency; Gemini 3.1 Pro is 7.5 times cheaper on input tokens than Claude Opus 4.6. For a developer running 100 million tokens through an agentic pipeline, this translates to a cost of $200 versus $1,500. The release introduces a new three-tier "thinking system," allowing developers to modulate the model's compute between "Low," "Medium," and "High" levels. This provides a trade-off between output latency and reasoning depth, optimizing for specific task complexities. Google is integrating 3.1 Pro into its agent-first development platform, Google Antigravity, to enhance capabilities in robust planning and navigating long-horizon tasks across entire codebases. The model also gains the ability to natively generate and animate Scalable Vector Graphics (SVGs) directly from text prompts.
