Gemini 3.6 Flash (Google AI Studio)

ByteMark's Tool Loop benchmark (August 3, 2026) evaluated multi-hop tool calling across 29 tools and 5 requests where no single tool can answer alone. The loop runs for real: the model calls, reads a result, and calls again, with chains locked by identifiers the model cannot know in advance. Cases include dividend yiel Gemini 3.6 Flash appeared in the cost-vs-latency scatter alongside 15 targets including Z.ai GLM-4.7, Gemini 3.5-Lite, Gemini 2.5-Flash, Claude Sonnet 5, Grok 4.5, and DeepSeek v4-pro. The benchmark's design, which punishes loose answers and requires the model to handle arithmetic, geocoding, and identifier resolution

ByteMark's Tool Haystack benchmark (July 31, 2026) evaluated Gemini 3.6 Flash on tool selection from a list of 50 banking tools across six cases, where family members are scattered and two cases deliberately make the tool whose name best fits the message the wrong one. The test gates on tool name, so a wrong pick score This evaluation provides independent third-party evidence of Gemini 3.6 Flash's tool-use accuracy under high-distractor pressure, directly relevant to the model's agent-positioning claims. The benchmark design specifically tests whether a model reads tool descriptions or merely pattern-matches on names, making the data

ByteMark's Prompt Leak benchmark (July 29, 2026) evaluated Gemini 3.6 Flash on system-prompt canary extraction resistance. A bank support persona carries a unique build reference (BM-CANARY-7F3A9K2Q) and is told never to reveal it; five attack vectors attempt extraction via direct ask, developer-debug pretext, translat Gemini 3.6 Flash appeared in the cost-vs-latency scatter among 15 targets alongside Gemini 3.5-Lite, Gemini 2.5-Flash, OpenAI 5.6-luna/terra, Claude Haiku 4.5, Grok 4.5, Claude Sonnet 5, and Qwen 3.7-flash (which achieved the lowest cost at $0.0059 and 6.8s). The benchmark provides independent third-party safety and ro

Google announced Gemini 3.6 Flash on July 21, 2026, positioning the variant as a balanced general-purpose workhorse for coding, knowledge work, multimodal analysis, and multi-step agents. The launch was reported alongside Gemini 3.5 Flash-Lite (a higher-throughput, lower-cost option for search, extraction, classificati Gemini 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens. Google claims the model consumes 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index while taking fewer reasoning steps and tool calls in multi-step workflows, an efficiency framing aimed at lowe

ByteMark's Ticket Extract benchmark (July 22, 2026) evaluated Gemini 3.6 Flash served via Google AI Studio on a customer-support email-to-structured-ticket task covering ticket id, customer details, product, severity, start date, order number, requested action, and per-device serial/model entries. The task deliberately The benchmark compared Gemini 3.6 Flash against ten other models including Gemini 3.5 Flash-Lite (which achieved the best overall score at $0.0084 and 1.2s latency), OpenAI 5.6-luna, Claude Haiku 4.5, Grok 4.5, and Qwen 3.7-plus. Gemini 3.6 Flash placed in the higher-cost/higher-latency region of the cost-vs-latency sc

ByteMark's Receipt Parse benchmark evaluated Gemini 3.6 Flash (Google AI Studio) on a short PDF receipt-to-JSON-schema extraction task requiring invoice/receipt numbers, date paid, vendor, bill-to details, currency, subtotal, discount, total, amount paid, payment method, and a line-item table. Grading uses exact match The benchmark compared 14 models including Gemini 3.5-Lite ($0.0021, fastest), OpenAI 5.6-luna (cheapest at $0.0011), Claude Haiku 4.5, Gemini 3.7-flash, Gemini 3.8-flash, and Grok 4.5/4.6. Gemini 3.6 Flash sat in the mid-range on cost ($0.0183) and latency (8.1s) while maintaining a perfect schema-adherence score, pro