Version 5

Google launched Gemini 3.5 Flash at Google I/O 2026 on May 19, priced at $1.50/M input and $9/M output tokens — 3x the cost of Gemini 3 Flash Preview and up to 6x the cheapest prior Flash option [12][13]. The model replaced Gemini 3 as the web default and powers Google Search AI Mode [6][1]. Google's counter-narrative frames the price as cheap versus GPT-5 and Claude, claiming over $1 billion in annual enterprise savings versus competing providers [20][21] — a framing that conflicts with developers' lived experience of a sharp internal price hike. A parallel billing-anomaly story has expanded: unexplained cost increases dating to at least March 2026 [30], a €54k bill from a single unrestricted Firebase key [31], and reports of price hikes that never appeared on the pricing page [32] suggest Gemini API billing behavior is murkier than the sticker price debate implies.

The pricing debate has split into two factually valid but opposite framings — cheap relative to OpenAI and Anthropic flagships, or expensive relative to prior Flash models — and which frame enterprise buyers adopt will define Gemini 3.5 Flash's market position. The expanding billing-anomaly story raises a sharper question: whether the cost of using the Gemini API is predictable at all, making billing reliability a more fundamental concern than sticker price.

Google launched Gemini 3.5 Flash at Google I/O 2026 on May 19 under the tagline 'frontier intelligence with action' [1][2]. The model had been spotted in the Google Cloud Console hours before the official announcement [3][4], alongside a companion model called Gemini Omni Flash [5]. At launch, Gemini 3.5 Flash replaced Gemini 3 as the default on Google's web interface and began powering AI Mode in Google Search [6]. Official documentation — including 'What's New' API guides [7][8], a Google Cloud Enterprise Agent Platform listing [9], and Google DeepMind's model card [10] — formalized its capabilities. The model was also integrated into updated Google AI subscription tiers announced at I/O [11].

The pricing landed at $1.50 per million input tokens and $9 per million output tokens — 3x the cost of Gemini 3 Flash Preview and 6x the cheapest prior Flash-tier option [12][13]. One analysis noted the output token price is nearly 30x more expensive than the previous cheapest Flash alternative [14]. An independent Artificial Analysis benchmark run cost $1,551.60 on Gemini 3.5 Flash versus $892.28 for Gemini 3.1 Pro Preview [12], nearly erasing the historical price gap between Flash and Pro tiers. On capability, Artificial Analysis rated Gemini 3.5 Flash as the 'clear leader on the Intelligence vs Speed Pareto frontier' [15][16], and an early-access user described first impressions as feeling 'on par with Gemini 3.1 Pro Preview' [17]. A Towards AI piece testing the model on 18 agent tasks found it outperformed GPT-5.5 on agentic workloads at 4x the speed, while framing its cost as 6x higher than prior Flash-Lite baselines [18]. One observer summarized the situation tersely: 'Gemini 3.5 Flash beats last year's Pro. The pricing conversation tells the rest.' [19]

Google and aligned voices have pressed a cross-vendor reframing: Gemini 3.5 Flash is cheap relative to GPT-5 and Claude flagships, not expensive relative to prior Flash models. VentureBeat reported Google's claim that the model can slash enterprise AI costs by more than $1 billion a year versus competing providers [20], and AI Business echoed the enterprise cost-efficiency angle [21]. An independent benchmark headline claimed the model scores within two points of Anthropic's flagship at a third of that flagship's price [22]. This reframing has not defused developer frustration: Reddit threads, social media posts, and developer forums continued characterizing the 3x internal price hike as a betrayal of the Flash-tier value proposition [23][24][25], with at least one developer explicitly calling for lower pricing [26]. Analyst Simon Willison documented the cost jump with benchmark data and framed it as part of an industry-wide repricing trend. Venture capitalist Tomasz Tunguz published 'The Unsustainable Subsidy' [27], arguing that current AI pricing reflects an ending land-grab subsidy rather than sustainable economics — lending independent analyst credibility to the thesis that the aggressive discounting era is closing across all major labs.

The billing-anomaly dimension of this story has expanded significantly and now predates the Gemini 3.5 Flash launch itself. A Google AI Developers Forum thread documented unexpected cost spikes on gemini-3-flash-preview despite decreased usage [28][29]. A separate 'URGENT' forum post reveals API cost increases dating to March 16-17, 2026 [30] — nearly two months before Google I/O — suggesting billing behavior changes were underway independently of the new model rollout. A Hacker News thread documented a €54k billing spike in 13 hours from an unrestricted Firebase browser key accessing the Gemini API [31], introducing a security-exposure dimension: predictable billing may require API key access controls that Google has not made sufficiently visible to developers. A Medium piece described a price hike that 'never showed up on the pricing page' yet caused bills to rise 27% [32]. Taken together, these reports suggest that Gemini API billing behavior — its transparency, security assumptions, and change-notification practices — is a distinct and potentially more consequential problem than the sticker-price debate. On tooling, the llm CLI plugin added gemini-3.5-flash support in version 0.32 [33] with a concurrent alpha adding streaming reasoning token support [34]; the llm-gemini GitHub repository [35] serves as the canonical integration point for CLI users. Gemini 3.5 Pro remains unannounced; one social media observer on May 23 claimed it is coming next month [36], but Google has made no formal statement.