Google DeepMind's latest AI model, Gemini 3.5 Flash, raises a pressing question for developers: how will increased costs impact the deployment of AI agents? With operational expenses soaring to more than five times those of its predecessor, Gemini 3 Flash, the new model's pricing structure signals a significant shift in the AI sector.
The substantial cost increase arises from a notable rise in token consumption during agent tasks. An analysis from Artificial Analysis indicates that running Gemini 3.5 Flash costs approximately 5.5 times more than the previous version and nearly doubles the expenses of the Pro model, Gemini 3.1. Although the context window remains at one million tokens, the price per token has tripled, with Google now charging $1.50 for input tokens and $9.00 for output tokens, up from $0.50 and $3.00, respectively.
While the per-token cost is still lower than that of the Gemini 3.1 Pro, the overall expenses present a different picture. Developers face a staggering 75% increase in costs compared to the Gemini 3.1 Pro due to the heightened token consumption associated with Gemini 3.5 Flash. This trend aligns with a broader industry pattern, as competitors like Anthropic and OpenAI have also raised their model prices. For example, OpenAI's GPT 5.5 saw an increase of 50% to 90% over its predecessor, GPT 5.4, while Anthropic's Opus 4.7 experienced a hidden price increase of 30% to 40%.
The Shift Towards Efficiency
As these models grow increasingly costly, the raw price of tokens is becoming a less valuable metric for developers and companies. What matters now is the efficiency of these models—specifically, how many tokens they consume to complete a given task. This shift underscores the necessity for AI providers to prioritize efficiency alongside performance.
Gemini 3.5 Flash demonstrates improved performance, scoring 55 on the Artificial Analysis Intelligence Index—nine points higher than its predecessor. It surpasses other models like Grok 4.3 and Claude Sonnet 4.6, showing improvements across nearly all tested categories. However, benchmark scores can be misleading; real-world performance will only be fully understood through extended usage across diverse tasks.
On the AA Omniscience index, which evaluates knowledge accuracy and hallucination tendencies, Gemini 3.5 Flash shows an improvement of 11 points, reducing its hallucination rate to 61%. This is a significant drop from its predecessor but still lags behind leaders such as MiMo-V2.5-Pro and Grok 4.3, which maintain hallucination rates of just 25%. This indicates that while the new model is smarter, challenges with inaccuracies remain.
Implications for the AI Market
The introduction of Gemini 3.5 Flash at these elevated costs raises critical questions about the future of AI deployment. Companies must now weigh the benefits of advanced AI capabilities against the financial implications of their usage. As token consumption continues to climb, the industry may need to adapt its pricing structures and models to align with efficiency metrics rather than just raw token costs.
As AI technology evolves, the increasing operational costs of these models highlight a crucial turning point for developers and businesses. The focus will likely shift toward balancing performance, cost, and efficiency to navigate this new terrain effectively. The implications of this trend could redefine how AI applications are developed and integrated into various sectors, marking a new era in the AI sector.
Quick answers
What is the main cost increase associated with Gemini 3.5 Flash?
Gemini 3.5 Flash costs over five times more to run than its predecessor due to high token consumption.
How does the token pricing compare to previous models?
Google has tripled the token prices, charging $1.50 for input and $9.00 for output tokens.
What are the performance improvements of Gemini 3.5 Flash?
Gemini 3.5 Flash scores 55 on the Artificial Analysis Intelligence Index, improving by nine points over Gemini 3 Flash.
How do hallucination rates compare across models?
Gemini 3.5 Flash has a hallucination rate of 61%, down 31 percentage points from its predecessor, but still higher than leading models.



