Gemini 2.5 Ultra Arrives with 2M Token Context and Native 60fps Video
Google DeepMind has released Gemini 2.5 Ultra, extending context to two million tokens and adding native video understanding at 60 fps. The model is available via Vertex AI and Google AI Studio and leads several long-context benchmarks.
Original sourceGoogle DeepMind has released Gemini 2.5 Ultra, the latest in its Gemini series, featuring a two-million token context window and native video processing at 60 frames per second. The model is accessible through Vertex AI for enterprise users and Google AI Studio for developers, continuing Google's pattern of dual-track distribution for its frontier models.
The two-million token context window is the headline number — roughly four times the previous Gemini 1.5 Pro ceiling and enough to ingest multiple book-length documents, full codebases, or hours of video in a single prompt. Google claims the model tops several long-context benchmarks, though the specific evaluations and methodologies have not yet been independently verified. Native 60fps video understanding is a meaningful architectural claim, suggesting the model processes temporal motion data rather than sampling sparse keyframes.
Access is currently available through existing Vertex AI and Google AI Studio accounts, with pricing details published in Google's standard API tiers. The release positions Gemini 2.5 Ultra directly against Anthropic's Claude models and OpenAI's GPT-4o in the long-context enterprise segment, a market where Google already has distribution advantages through its cloud infrastructure and Workspace integrations.
The practical implications for developers center on workflows that previously required chunking or summarization pipelines — legal document review, large codebase analysis, and long-form video summarization being the most immediate use cases. Whether the model maintains coherent reasoning across the full two-million token window, rather than just technically accepting that many tokens, remains the key question practitioners will stress-test in the coming weeks.
Panel Takes
The Builder
Developer Perspective
“The primitive here is a long-context inference endpoint — and the real test is whether it actually reasons coherently at 1.8M tokens or just technically accepts the input and degrades. Two million tokens sounds clean until you're paying for prefill on a 400K-token codebase and the model hallucinates a function that was clearly defined on page three. I want to see latency numbers, cost-per-million-token benchmarks at scale, and whether the Vertex AI SDK handles streaming on contexts this size without timing out before I get excited about the context ceiling.”
The Skeptic
Reality Check
“'Tops several long-context benchmarks' is doing a lot of work in this announcement — Google authored or selected those benchmarks, and independent evals on multi-million token coherence are notoriously easy to game by optimizing for specific retrieval patterns. The 60fps video claim is interesting but unverified; the difference between native temporal understanding and fast keyframe sampling is real and won't be settled by a blog post. What kills this in 12 months isn't a competitor — it's Google itself shipping a cheaper, faster 2.5 Flash that handles 90% of these use cases at a fraction of the cost, making Ultra a prestige product with thin adoption.”
The Futurist
Big Picture
“The bet embedded in a 2M token context window is specific and falsifiable: that the cost of attention over massive contexts will fall faster than the value of chunking pipelines rises, making long-context a default rather than a premium. The second-order effect nobody is talking about is what happens to the entire RAG tooling ecosystem — if you can fit a full knowledge base in context reliably, the retrieval layer becomes optional overhead, and a significant chunk of the LangChain/LlamaIndex stack becomes redundant. Google is riding the inference cost curve and betting it bottoms out before developers build irreversible habits around retrieval-augmented architectures.”
The Founder
Business & Market
“The moat here isn't the model — it's distribution through Vertex AI, where Google already has procurement relationships with every enterprise that runs on GCP, meaning Gemini 2.5 Ultra gets evaluated without a separate sales motion. The pricing architecture will determine whether this is a real business line or a loss-leader to lock workloads into Google Cloud; if token costs at 2M context are priced at marginal cost, the enterprise legal and media use cases Google is targeting won't survive a unit economics check. The business survives a smarter competitor easily — it doesn't survive Google treating the model as a cloud attachment rather than a standalone revenue line.”