Google Ships Three New Gemini Flash Models While Its Flagship 3.5 Pro Stays in the Lab

Google used its latest Gemini refresh to prioritize agents over headline benchmarks, shipping Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber while leaving its long-awaited Gemini 3.5 Pro model in partner-only testing. The company has not given a firm release window for 3.5 Pro, saying only that it will ship once it clears internal readiness checks.

The centerpiece of the release is Gemini 3.6 Flash, which Google is positioning as an agentic workhorse rather than a raw reasoning champion. On OSWorld-Verified, a benchmark that measures how well a model can operate a computer interface autonomously, the model scored 83.0%, up from 78.4% posted by Gemini 3.5 Flash. It also climbed to 63.9% on MLE Bench, a machine-learning engineering test, compared with 49.7% for its predecessor, and improved to a 49% success rate on DeepSWE, a software-engineering benchmark, from 37% previously.

Google is pairing those gains with an efficiency pitch aimed squarely at enterprise budgets. According to figures from the Artificial Analysis Index, Gemini 3.6 Flash consumes 17% fewer output tokens than the 3.5 Flash generation while completing similar tasks, a detail Google is using to argue that agentic workloads, which often run in long, iterative loops, will get materially cheaper to operate at scale. Separately, some reporting has cited a 65% reduction in token usage specifically on the Datacurve DeepSWE benchmark, a steeper figure than the general 17% efficiency claim, suggesting the larger savings may be workload-specific rather than representative of average usage.

Pricing for the new tier has also been reported with some inconsistency. Gemini 3.6 Flash is listed at $1.50 per million input tokens and $7.50 per million output tokens, while Gemini 3.5 Flash-Lite comes in considerably cheaper at $0.30 per million input tokens and $2.50 per million output tokens. Flash-Lite is also billed as the speed option, running at 350 output tokens per second on the Artificial Analysis Index, the fastest figure among the 3.5-series models, positioning it as the low-latency, low-cost choice for high-volume, less complex tasks.

Google has also folded client-side computer-use tooling directly into the Gemini API and the Gemini Enterprise platform, letting developers build agents that interact with software interfaces without stitching together separate automation layers. Legal-tech platform Harvey and research tool Hebbia are named as early adopters, both using Gemini 3.6 Flash for multimodal document processing, a use case that plays to the model's combined gains in coding, knowledge work, and multimodal understanding.

The absence of Gemini 3.5 Pro from this rollout is the notable gap. Google has kept the model restricted to select partners for testing, a signal that the company is treating its next flagship reasoning model as a separate, slower-moving release track from the faster iteration cycle now visible in the Flash lineup.