Gemini 2.5 Ultra

Gemini 2.5 Ultra

Note: "Gemini 2.5 Ultra" is not a distinct model. "Ultra" refers to Google AI Ultra — Google's $249.99/month (later $99.99/month) subscription tier, not a separate model variant. The top Gemini 2.5 model is gemini-2-5-pro with its Deep Think reasoning mode, which is gated to Ultra subscribers. This page covers Gemini 2.5 Pro + Deep Think as the highest-capability 2.5-generation offering.

Gemini 2.5 Pro + Deep Think

Gemini 2.5 Pro is Google DeepMind's flagship 2.5-series model, released to general availability June 17, 2025. Deep Think is an extended reasoning mode for 2.5 Pro that runs parallel chains of thought — "generating many ideas at once, considering them simultaneously, and revising or combining them" before producing an answer. Deep Think was previewed at Google I/O May 2025 and rolled out to Ultra subscribers August 1, 2025.

The Gemini 2.5 generation has since been succeeded by Gemini 3.x (Gemini 3.1 Pro GA, Gemini 3.5 Flash announced at Google I/O May 2026). As of June 2026, gemini-2-5-pro is stable and still available on the API but is no longer the flagship.

Specs

Property Value
Context window 1,048,576 tokens (1M)
Max output 65,536 tokens
Deep Think output limit 192,000 tokens
Input modalities Text, images, audio, video

Benchmarks (Deep Think mode)

Benchmark Score
LiveCodeBench V6 87.6% (state-of-the-art at release)
Humanity's Last Exam Leading at release
2025 IMO Bronze-level performance
MMMU (multimodal) 84.0%

Standard 2.5 Pro led LMArena and the WebDev Arena (ELO 1415) at its May 2025 launch.

Pricing (Gemini 2.5 Pro)

Tier Input Output
≤200K tokens $1.25/MTok $10/MTok
>200K tokens $2.50/MTok $15/MTok

Deep Think is not separately priced for API use; access was gated to trusted testers via the Gemini API as of August 2025.

Access

  • Consumer: Google AI Ultra subscription ($99.99/month as of 2026, reduced from $249.99). Deep Think toggled in the Gemini app prompt bar; limited daily quota.
  • Developer: Google AI Studio (free tier, rate-limited). Vertex AI (paid). Deep Think in API as of late 2025 via select tester program.

Weaknesses

  • Context window is 1M tokens — gemini-2-5-pro is a better reference for detailed capability notes.
  • Deep Think has a daily usage cap for consumer subscribers.
  • Gemini 3.x generation (Gemini 3.1 Pro) now leads on most benchmarks.

Related

gemini-2-5-pro · gemini-3-5-flash · extended-thinking · evals · rag

Sources