Google Broadens Availability of Gemini 2.5 Pro Following Impressive Benchmark Performance
Overview of Google’s Gemini 2.5 Pro
Google has recently made the Gemini 2.5 Pro, its latest AI model, more accessible to the public. This AI tool has shown impressive results in scientific testing and is now offered at competitive prices. Alphabet’s CEO, Sundar Pichai, highlighted that Gemini 2.5 Pro is not only Google’s most intelligent AI model but also the most sought-after. In fact, there has been an 80% increase in demand across various platforms like Google AI Studio and the Gemini API this month.
Access and Usage
Starting this week, users can access a broader public preview of Gemini 2.5 Pro, which offers higher usage limits. A free tier is available for those interested. Users of Gemini Web Chat are able to use the 2.5 Pro Experimental model, which offers similar performance capabilities. Additional announcements regarding Gemini’s features and updates are expected during the upcoming Cloud Next ’25 conference on April 9.
Competitive Pricing Structure
Gemini 2.5 Pro employs a tiered pricing model for its API. For prompts with up to 200,000 tokens, the input cost is $1.25 per million tokens, while output will cost $10 per million tokens. If the prompts exceed this limit, the costs increase to $2.50 for input and $15 for output per million tokens. Though prompt caching is not currently available, Google may introduce this feature in the future, potentially lowering overall costs.
Users can also benefit from Google’s free grounding with Google Search, which allows up to 500 queries daily, followed by an additional 1,500 free queries. After reaching this limit, users will be charged $35 for each 1,000 queries. Notably, data from the free tier might be used for AI training, whereas data from the paid tier will remain private.
Performance in Benchmarking
Gemini 2.5 Pro has garnered attention for its outstanding performance in scientific benchmarks. According to EpochAI, this model achieved a score of 84% on the GPQA Diamond benchmark, significantly outperforming the usual 70% score achieved by human experts. This benchmark consists of complex multiple-choice questions in fields like biology, chemistry, and physics.
Furthermore, Gemini 2.5 Pro scored an impressive 18.8% on "Humanity’s Last Exam," the highest among models that were tested without additional tools. This performance highlights its advanced reasoning capabilities, although Google has not released detailed information on the model’s architecture or the technical requirements used during its training.
Cognitive Abilities and Feedback
In other forms of testing, such as a weekly assessment conducted by TrackingAI.org, Gemini 2.5 Pro showcased strong cognitive abilities with an average IQ score of 130, well above the typical range for AI models. These IQ tests utilized a text-based version of the Norwegian Mensa IQ test, ensuring that the questions were not in the training data. If the model hesitates to provide an answer, it is given multiple chances to respond before its final answer is recorded.
Prominent figures in the AI and tech fields have expressed positive opinions about Gemini 2.5 Pro. Notably, computer scientist François Chollet has reported that this model is his go-to option for most tasks, excluding those related to image generation. Investor Martin Casado and researcher Shane Gu have also praised the model, specifically highlighting its cost-efficiency for various programming and research tasks.
Overall, with its robust capabilities, competitive pricing, and increasing accessibility, Google’s Gemini 2.5 Pro stands poised to make a significant impact on the AI landscape.
