# Google’s Latest AI Model Arrives with a Massive 1 Million Token Output Window
## What Is Google’s Newest Frontier Model?
Google DeepMind has opened the doors to its latest large language model, making it available to developers, businesses, and paying subscribers. This model belongs to a family known for combining multiple specialized “experts” inside a single system, allowing it to handle complex, multi-step tasks efficiently. It is designed with long workflows in mind, particularly in areas like software development, financial analysis, legal research, scientific discovery, and digital security.
Unlike smaller models that struggle with extended conversations or lengthy documents, this new system can generate responses up to one million tokens in a single go. That is a dramatic increase from the 64,000-token limit found in earlier versions, opening the door to everything from full-length code generation to comprehensive legal document analysis in one session.
The model accepts text, images, audio, and video as input, and produces text-based output. Google built it using custom TPU hardware and a proprietary training framework, keeping the weights closed and available only through its cloud platform.
## How Does It Perform in the Real World?
Google has shared a range of benchmark results to illustrate where the model excels and where it still has room to grow.
### Where It Leads the Pack
The model achieves the top score in several key evaluations. On a test measuring long-form video understanding, it reached 91.7%, setting a new high point for Google’s lineup. It also dominates in a knowledge-work benchmark at 68.9%, and shows strong results on automation tasks, scoring 51.3% compared to 42.5% from the closest competitor.
In software engineering tasks, it posted 77.9% on a deep coding evaluation, ahead of rival systems that scored in the low-to-mid 70s. On an embodied reasoning benchmark designed to test physical-world problem solving, it reached 87%.
### Where It Falls Behind
The model is not without weaknesses. On several coding and operating system benchmarks, it trails behind competitors. In one software engineering test, it scored 55% while a rival hit 65.5%. On a terminal-based task benchmark, it scored 57.4% compared to a competitor’s 66.4%. It also underperforms on a post-training evaluation at 45.3% versus a rival’s 49.3%.
Google has also flagged occasional hallucinations, slowdowns, and timeouts in the model card. Under its internal safety framework, the company acknowledges it cannot fully rule out certain advanced capabilities related to biology and cybersecurity, though it says it has deployed safeguards to reduce those risks.
## How Much Does It Cost?
Google is offering introductory pricing of $2 for every million input tokens and $10 for every million output tokens. This introductory rate is valid until November 30, 2026. After that date, prices rise to $4 per million for input and $20 per million for output.
To put that in perspective, the new model is significantly cheaper than some competitors. One rival charges $10 for input and $50 for output per million tokens, while another charges $4 and $20. Cached input tokens receive a 95% discount, which can meaningfully reduce costs for applications that reuse the same prompts repeatedly.
## Where Can You Access It?
The model is available exclusively through Google’s cloud ecosystem. Developers can tap into it using the Gemini API, Google AI Studio, and Android Studio. Enterprise customers have additional pathways through the Gemini platform, Gemini for business, and Google’s Antigravity offering.
Everyday consumers who subscribe to Google’s Pro or Ultra plans can use it through the Gemini app and the AI Mode feature in Google Search. Google also made the model available to a select group of cybersecurity professionals before the public launch, providing access without certain safety guardrails for defensive research purposes.
## How Does It Stack Up Against the Competition?
| Feature | Google’s Latest Model | Rival A | Rival B | Rival C |
|—|—|—|—|—|
| Architecture | Sparse mixture of experts | Not disclosed | Not disclosed | Not disclosed |
| Total context window | 1 million tokens | 1,050,000 | 1 million | 1 million |
| Maximum output | Up to 1 million tokens | 128,000 | 128,000 | 128,000 |
| Input types | Text, images, audio, video | Text, images | Text, images | Text, images |
| Introductory price (in/out per 1M) | $2 / $10 | $10 / $50 | $4 / $20 | $10 / $50 |
| Standard price (after Nov 30) | $4 / $20 | $10 / $50 | $4 / $20 | $10 / $50 |
The standout advantage here is the combination of a massive output window and a competitive price point. Most rivals cap output at 128,000 tokens, while this model allows ten times that in a single response. The pricing, especially during the introductory period, undercuts several competing offerings.
—
## Frequently Asked Questions
**What kind of model architecture does this system use?**
It uses a sparse mixture-of-experts design, which means only a small subset of the model’s internal networks activate for any given token. This makes it more efficient than a dense model of comparable size, though Google has not disclosed the total number of parameters.
**Can I run this model on my own hardware?**
No. The model is only available through Google’s cloud services and cannot be downloaded or self-hosted. There are no open weights available.
**What happens when my prompt uses most of the 1 million token window?**
Since the input and output share a single 1 million token budget, a longer prompt leaves less room for the model’s response. Google has published documentation on a continuation protocol to help manage very long outputs.
**Is the model available worldwide?**
Access is available through Google’s standard cloud regions and services, though specific availability details depend on your subscription tier and geographic location.
**Does Google claim this is the most capable model in every area?**
Google’s own benchmarks show the model leading in several categories, but independent evaluations tell a more nuanced story. It ranks last among comparable systems on certain coding and operating system benchmarks, so it is not universally superior.
**What are the known limitations?**
Users should expect occasional hallucinations, delayed responses, and timeouts during heavy use. The model’s knowledge has a cutoff date of March 2026, meaning it lacks awareness of more recent events.
—
## Conclusion
Google’s newest model arrives with a bold promise: the ability to handle extraordinarily long tasks in a single response, all at a competitive price. Its mixture-of-experts architecture and 1 million token output window set it apart from most offerings currently available on the market. It performs exceptionally well on video understanding, knowledge work, and automation benchmarks.
However, it is not without shortcomings. It trails competitors on several software engineering and operating system tasks, and Google itself has acknowledged issues with hallucinations and occasional performance hiccups. For developers and enterprises weighing whether to adopt it, the decision will depend heavily on the specific use case, particularly whether the long output capability and lower pricing align with their needs.
As with any frontier model, it is worth testing carefully before committing to production workloads, keeping in mind that benchmark scores represent just one dimension of real-world usefulness.
Thank you for reading
