Google has launched Gemini 3.1 Ultra, its most significant model release of the year. The new model represents Google DeepMind’s most aggressive push yet against OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6.
What’s New in Ultra
- 2 million token context window — works natively across text, image, audio, and video without transcription intermediaries.
- Native multimodal reasoning — the model was trained from scratch to reason across all modalities simultaneously, not via post-hoc routing.
- Sandboxed Code Execution tool — Ultra can write, run, and test code mid-conversation inside a sandboxed environment.
- Agentic tool chaining — Ultra can plan and execute multi-step workflows across web, code, and data tools without user prompting.
Benchmark Results
Gemini 3.1 Ultra tops key reasoning benchmarks. On GPQA Diamond, Ultra scores 94.3 percent. On OSWorld-Verified (computer-use), Ultra hits a record 68.2 percent. On OpenAI’s GDPval knowledge-work test, Ultra scores 85 percent, overtaking GPT-5.4’s March score of 83 percent.
Pricing
Google is pricing Ultra at $8 per million input tokens and $24 per million output tokens — positioned above Gemini 3.1 Pro ($1.25/$5) but below Claude Opus 4.6 ($15/$75). Developers on the Gemini API waitlist can begin using the model today; consumer rollout through Gemini Advanced starts next week.
Why It Matters
Ultra lands against a backdrop of Google’s Apple partnership announcement — the Siri-Gemini integration debuts in 2026 and will eventually run on 1.4 billion iPhones via Apple’s Private Cloud Compute architecture. Ultra is reportedly the same model family Apple will route queries to. Combined with Apple’s decision, today’s launch cements Google as the unambiguous 2026 model leader — at least until GPT-6 and Claude 4.7 arrive.
Follow Vibes Uncut Media for continuing coverage of the AI model race.
Leave a Reply