Google releases trio of Gemini models, but Gemini 3.5 Pro still delays

Google introduced three fresh Gemini models this week: Gemini 3.6 Flash, a budget variant called 3.5 Flash Lite, and 3.5 Flash Cyber tailored for vulnerability detection. The launches come with a notable gap—Gemini 3.5 Pro, Google's flagship offering originally promised for June following its May I/O announcement, remains unavailable to the broader public and is still undergoing partner testing.

Flash 3.6 brings speed gains at lower cost#

The Gemini 3.6 Flash upgrade delivers meaningful improvements over its predecessor while trimming expenses. Input pricing holds steady at $1.50 per million tokens, though output costs drop from $9 to $7.50—a 17% token reduction in agentic workflows, according to Artificial Analysis data.

Performance jumps are particularly visible in software engineering tasks. On the DeepSWE benchmark, 3.6 Flash reaches 49% compared to 3.5 Flash's 37%, though competitors like Claude Sonnet 5 edge ahead at 54%. Similar gains appear across machine learning research and computer use benchmarks, positioning it competitively as a mid-range offering.

On general knowledge work, the model scores 1421 on GDPval-AA v2—a notable leap from 3.5 Flash's 1349—yet still trails competing models from other labs. The pricing calculus remains complex; GPT 5.6 Luna, for instance, undercuts Gemini on both speed and cost depending on context window size.

Lite model targets high-throughput workloads#

Google's 3.5 Flash Lite version aims at speed-critical tasks requiring minimal reasoning, such as document processing and search. It costs $0.30 for input and $2.50 for output per million tokens—up from the prior lite model's $0.25/$1.50. Performance metrics show substantial gains over its predecessor, with GDPval-AA v2 jumping from 642 to 1140 and software engineering benchmarks rising from 49.6% to 54.2%. Companies processing high volumes of data with tight latency constraints find the trade-offs appealing; one enterprise noted it hits "the Pareto frontier" for accuracy, speed, and expense in receipt extraction workflows.

Cybersecurity model arrives with limits#

The 3.5 Flash Cyber variant applies Gemini's capabilities to finding software vulnerabilities, using identical pricing to the standard Flash model. Google says it matches the performance level of Anthropic's Claude Opus 4.6. However, access remains tightly controlled—restricted to governments and select partners through Google's CodeMender tool, with broader availability planned only gradually. This cautious rollout reflects industry anxiety about AI-powered security exploitation tools following regulatory scrutiny of earlier releases.