Google Unveils Gemini 2.5 Ultra With 1M Tokens

Google DeepMind has officially launched Gemini 2.5 Ultra, the company's most capable large language model to date, featuring a groundbreaking 1-million-token context window and state-of-the-art performance across virtually every major AI benchmark. The model represents Google's strongest bid yet to reclaim the frontier AI crown from rivals like OpenAI, Anthropic, and Meta.

Available initially through Google AI Studio and the Gemini API, the 2.5 Ultra model targets enterprise developers and researchers who need the deepest reasoning capabilities and the ability to process massive documents, codebases, and multimedia inputs in a single prompt. Google is also rolling out access to Gemini Advanced subscribers on a phased basis.

Key Facts at a Glance

• Context window: 1 million tokens — equivalent to roughly 700,000 words or 1,500 pages of text
• Benchmark dominance: Tops leaderboards on MMLU-Pro, GPQA Diamond, HumanEval, and multiple multimodal reasoning tests
• Modality support: Natively handles text, images, audio, video, and code in a single unified model
• Pricing: API access starts at $7 per million input tokens and $21 per million output tokens
• Availability: Launching first in the US and EU, with broader global rollout over the coming weeks
• 'Thinking' mode: Built-in chain-of-thought reasoning that can be toggled on for complex tasks

What Makes Gemini 2.5 Ultra Different From Previous Versions

Gemini 2.5 Ultra is not simply an incremental upgrade over its predecessor, Gemini 1.5 Ultra. Google DeepMind describes it as a 'from-scratch' architecture revision that fundamentally rethinks how the model handles long-context reasoning, multimodal integration, and agentic task execution.

The most immediately striking feature is the 1-million-token context window. While Gemini 1.5 Pro previously offered a similar context length, this is the first time Google has paired that capacity with its highest-tier 'Ultra' model. The result is a system that can ingest an entire novel, a full-length movie transcript, or a sprawling enterprise codebase — and reason over it coherently.

Unlike OpenAI's GPT-4o, which offers a 128,000-token context window, or Anthropic's Claude 3.5 Sonnet with 200,000 tokens, Gemini 2.5 Ultra's context length is roughly 5 to 8 times larger. This difference matters enormously for enterprise use cases like legal document analysis, financial modeling across thousands of pages, and full-repository code review.

Benchmark Performance Sets a New Standard

Google claims Gemini 2.5 Ultra achieves the highest scores ever recorded on several prominent AI benchmarks. On MMLU-Pro, the enhanced version of the Massive Multitask Language Understanding test, the model reportedly scores above 90%, surpassing both GPT-4o and Claude 3.5 Opus.

The model also excels in specialized reasoning domains:

• GPQA Diamond (graduate-level science questions): Gemini 2.5 Ultra leads all publicly reported models
• HumanEval (code generation): Achieves a pass rate above 92%, rivaling dedicated coding models
• MathVista (visual math reasoning): Outperforms all competitors in tasks requiring chart and graph interpretation
• Video-MME (video understanding): Sets a new state of the art for long-form video comprehension
• LMSYS Chatbot Arena: Early ELO ratings place it in the top tier alongside Claude 3.5 and GPT-4o

These results, if independently verified, would position Gemini 2.5 Ultra as the single most capable general-purpose AI model available today. However, independent benchmarking organizations have not yet completed their full evaluations, so these claims should be treated with appropriate caution.

The 'Thinking' Mode Changes How Users Interact With AI

One of the most significant additions in Gemini 2.5 Ultra is its native 'thinking' mode, which enables transparent chain-of-thought reasoning. When activated, the model explicitly shows its reasoning steps before arriving at a final answer — similar to the approach popularized by OpenAI's o1 and o3 series of reasoning models.

This feature is particularly valuable for complex tasks like multi-step mathematical proofs, legal analysis requiring citation chains, and debugging intricate software systems. Developers can toggle thinking mode on or off via the API, giving them fine-grained control over the trade-off between response speed and reasoning depth.

Google reports that thinking mode increases accuracy on hard reasoning benchmarks by 15-25% compared to the model's standard inference mode. The trade-off is latency — responses in thinking mode take approximately 2 to 4 times longer to generate, and they consume more tokens, which increases API costs.

Pricing and Access: How Gemini 2.5 Ultra Compares on Cost

Google has set Gemini 2.5 Ultra API pricing at $7 per million input tokens and $21 per million output tokens. This positions it as a premium offering, roughly comparable to OpenAI's GPT-4o pricing but significantly more expensive than lighter models like Claude 3.5 Sonnet or Google's own Gemini 2.5 Flash.

For context, here is how the pricing stacks up against major competitors:

• Gemini 2.5 Ultra: $7 input / $21 output per million tokens
• OpenAI GPT-4o: $5 input / $15 output per million tokens
• Anthropic Claude 3.5 Opus: $15 input / $75 output per million tokens
• Gemini 2.5 Flash: $0.15 input / $0.60 output per million tokens

The massive context window does create a unique value proposition, however. Organizations that previously needed to implement complex Retrieval-Augmented Generation (RAG) pipelines to work with large document sets may find that simply feeding everything into Gemini 2.5 Ultra's context window is both simpler and more accurate. This could reduce overall system complexity and engineering costs, even if the per-token price is higher.

Enterprise and Developer Implications Are Substantial

For enterprise customers, Gemini 2.5 Ultra opens several doors that were previously closed or required expensive workarounds. The million-token context window means that entire regulatory filings, patent portfolios, or multi-year financial reports can be analyzed in a single API call.

Developers building AI agents — autonomous systems that can plan, execute, and iterate on complex tasks — stand to benefit significantly. Google has emphasized that Gemini 2.5 Ultra is optimized for agentic workflows, with improved tool use, function calling, and structured output generation. The model can maintain coherent plans across extremely long interaction histories, which is critical for agents that need to operate over extended periods.

Google Cloud customers will also gain access to Gemini 2.5 Ultra through Vertex AI, Google's enterprise ML platform. This integration includes enterprise-grade features like data residency controls, VPC-SC support, and customer-managed encryption keys — features that matter deeply to regulated industries like healthcare, finance, and government.

The Competitive Landscape Intensifies

The launch of Gemini 2.5 Ultra comes at a particularly competitive moment in the AI industry. OpenAI is widely expected to release GPT-5 in the coming months, while Anthropic recently launched Claude 4 with enhanced reasoning capabilities. Meta continues to push the open-source frontier with its Llama series, and Chinese labs like DeepSeek have demonstrated that high performance is achievable at dramatically lower costs.

Google's strategy with Gemini 2.5 Ultra appears to be differentiation through integration. No other company can offer a frontier AI model that is simultaneously embedded in search (Google Search), productivity software (Google Workspace), cloud infrastructure (Google Cloud), mobile operating systems (Android), and consumer hardware (Pixel devices). This ecosystem advantage is difficult for pure-play AI companies like OpenAI or Anthropic to replicate.

The model also reflects Google's broader push toward multimodal AI as the default paradigm. Rather than treating text, images, video, and audio as separate capabilities bolted onto a language model, Gemini 2.5 Ultra processes all modalities natively within a single architecture. This unified approach could prove decisive as AI applications increasingly require simultaneous understanding of multiple data types.

Looking Ahead: What Comes Next for Google's AI Strategy

Google has signaled that Gemini 2.5 Ultra is just one piece of a larger puzzle. The company is expected to announce deeper integrations with Google Workspace at its upcoming Cloud Next conference, potentially enabling features like automatic meeting summarization from Google Meet recordings, intelligent document drafting in Google Docs, and advanced data analysis in Google Sheets — all powered by the 2.5 Ultra model.

There are also strong indications that Google is working on expanding the context window even further. Research papers from DeepMind have explored 10-million-token context lengths, suggesting that future Gemini versions could process entire libraries of information in a single prompt.

For developers and businesses evaluating their AI strategy, the key takeaway is clear: the frontier is advancing rapidly, and the gap between leading models is narrowing. Gemini 2.5 Ultra's combination of massive context, strong reasoning, multimodal capabilities, and deep ecosystem integration makes it a compelling option — but the choice between Google, OpenAI, and Anthropic increasingly depends on specific use cases, existing infrastructure, and long-term platform bets rather than raw model performance alone.

The AI arms race shows no signs of slowing down. With Gemini 2.5 Ultra, Google has made its strongest move yet — but the next volley from competitors is likely just weeks away.