Google DeepMind's Gemini Flash Series Expands: 3.6, 3.7 Flash, Flash-Lite and Flash Cyber — Pro Tier Still Delayed
Summary
- • Google DeepMind releases three new Gemini Flash models focused on agentic efficiency.
- • Gemini 3.6 Flash cuts output tokens 17% and costs less than 3.5 Flash.
- • Cybersecurity model Flash Cyber is limited to governments and trusted partners only.
- • Flagship 3.5 Pro remains delayed while Google begins its most ambitious Gemini 4 pre-training.
Updates
Gemini 3.8 Flash: third Flash release in six weeks at unchanged 3.7 Flash pricingrevised
Released Sep 2, 2026 at $0.75/1M input and $3.75/1M output tokens — same introductory price as 3.7 Flash. Delivers substantial gains over 3.7 Flash in software engineering, agentic tasks, and multi-step reasoning in specialized domains including quantitative finance and legal.
3.8 Flash tops DeepSWE v1.1 vs larger frontier models; HLE-Verified 54.9%
Outperforms most larger frontier models on DeepSWE v1.1 (long-horizon software engineering). Scores 54.9% on HLE-Verified (multi-step reasoning across STEM, humanities, professional fields). Leads Vals Finance Agent V2 and Harvey's Legal Agent Benchmark vs. 3.7 Flash and other frontier models.
3.8 Flash 'works harder': adaptive effort with extra reasoning steps and tool calls
On complex tasks, 3.8 Flash executes additional reasoning steps and iterative tool calls, potentially using more tokens to maximize performance. Developers can select lower effort levels to minimize token overhead, or continue using 3.7 Flash for efficiency-first workloads — both models remain fully supported.
3.8 Flash Cyber via new Fairwind Program: frontier vuln detection and automated patching
Successor to 3.5 Flash Cyber. Available to trusted defenders via the new Fairwind Program. Reaches frontier-level performance on CyberGym benchmark, surpassing both 3.5 Flash Cyber and larger frontier models. Internal benchmark: 70%+ vulnerability discovery success rate across 20 programming languages, with emphasis on automated patching over detection alone.
Agentic Video Understanding Launched Sept 1 Across Gemini Flash Models
Available across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Dynamically searches, scans, and inspects video segments (frames, audio, transcripts) rather than static 1 FPS ingestion. Launched via Gemini API in Google AI Studio and Gemini Enterprise Agent Platform at no additional cost.
Agentic Video: Up to 88% Fewer Tokens, 66% Lower Cost, 7% Better Accuracy
Compared to static video processing across standard benchmarks: token consumption reduced up to 88%, analysis costs reduced up to 66%, and accuracy improved up to 7%. Gemini 3.7 Flash with agentic video sits at the accuracy-to-cost Pareto frontier among tested models.
Active Goal-Directed Processing Replaces Fixed-Rate Video Ingestion
Instead of ingesting video at a fixed frame rate (default 1 FPS via API), Gemini takes an active role — determining what to watch, at what speed, and via which modality (frames, audio, or transcript), fetching only relevant moments. Especially effective for long-form video (10-min how-to guides to multi-hour recordings) where static processing forces cost-vs-detail trade-offs.
New Video Use Cases: Moment Retrieval, Anomaly Detection, Long-form Search, Counting
Sub-second moment retrieval for precise video editing; long-form needle-in-haystack queries across multi-hour videos without consuming millions of tokens; dynamic anomaly detection by resampling at higher FPS; and accurate counting of repeated actions or objects over time. Feature works on both uploaded videos and YouTube videos.
Gemini 3.7 Flash Released at Half the 3.6 Flash Cost
Released Aug 13, 2026 — three weeks after 3.6 Flash. Introductory price of $0.75/1M input and $3.75/1M output tokens, half of 3.6 Flash's original pricing. Positioned as Google's most intelligent workhorse model for coding and agents, incorporating developer feedback and algorithmic innovations.
3.7 Flash vs. 3.6 Flash Benchmark Comparisons
FrontierCode 1.1 Main: 43.6% vs 34.4%; DeepSWE v1.1: 65.3% vs 49.0%; WebDev Arena Elo: 1588 vs 1538; GDP.pdf (complex document reasoning): 34.0% vs 22.0%; AutomationBench (real-world business workflows): 30.4% vs 17.0%.
Gemini Spark Upgraded to 3.7 Flash Across 160+ Countries
Gemini Spark (Google's 24/7 personal AI agent, available to AI Pro and Ultra subscribers in 160+ countries) now runs on 3.7 Flash. Improves Workspace tool use, multi-skill workflow accuracy, and output quality for complex knowledge work tasks such as consolidating files, drafting emails, and updating documents.
Details
3.6 Flash Released
Improved workhorse model with better coding, knowledge work, and multimodal performance. 17% fewer output tokens vs. 3.5 Flash per Artificial Analysis Index; priced at $1.50/1M input and $7.50/1M output tokens — lower than 3.5 Flash.
3.5 Flash-Lite Released
Fastest, most cost-effective Gemini model at 350 output tokens/second per Artificial Analysis Index; designed for high-throughput agentic workflows and low-latency tasks at scale.
3.5 Flash Cyber in CodeMender
Specialized cybersecurity model bundled with the CodeMender code-security agent; trained to find and fix vulnerabilities. Exclusively available to governments and trusted partners in a limited-access pilot program.
3.6 Flash Benchmark Gains
Compared to 3.5 Flash: DeepSWE 49% vs. 37%; OSWorld-Verified 83.0% vs. 78.4%; MLE Bench 63.9% vs. 49.7%; GDPval-AA v2 1421 vs. 1349. Token reduction up to 65% in some tasks.
3.5 Pro Delay Confirmed
Bloomberg reported internal performance goal failures. Google's Logan Kilpatrick says 3.5 Pro is in partner testing and hopes to 'land soon.' Last Pro update was in February 2026, months behind rival labs.
Gemini 4 Pre-Training Launched
Google DeepMind has started its 'most ambitious pre-training run yet' for Gemini 4, running in parallel with current releases and indicating investment in the next model generation.
Enhanced Frontier Safety on 3.6 Flash
3.6 Flash ships with strengthened CBRN (Chemical, Biological, Radiological, Nuclear) and cyber-offense misuse safeguards, making it substantially more resistant to jailbreaks while minimizing false refusals for legitimate uses.
Competitive Lag on Flagship Tier
Since Google's last Pro update in February 2026, OpenAI released GPT-5.5 and began rolling out GPT-5.6; Anthropic launched Claude Opus 4.8, Claude Sonnet 5, and expanded frontier Fable 5 model access.
Each row covers a distinct development from Google DeepMind's July 21 Gemini Flash release announcement.
What This Means
Google DeepMind has shipped Gemini 3.8 Flash — its third Flash-tier model in six weeks — at the same price as 3.7 Flash, continuing a rapid iteration cadence that is aggressively compressing the capability-cost curve for production AI agents. The 'works harder' design approach lets 3.8 Flash trade token efficiency for significantly better performance on complex multi-step tasks in finance, legal, and long-horizon software engineering, while developers can tune effort levels to manage cost. Simultaneously, 3.8 Flash Cyber — available via the new Fairwind Program — achieves 70%+ vulnerability discovery across 20 programming languages with an emphasis on automated patching, building a dedicated cybersecurity AI track alongside the general-purpose Flash line. With six Flash models shipped since July 2026 and Gemini 4 pre-training underway, Google's Flash-tier strategy is firmly focused on winning the high-volume developer and enterprise market while its Pro tier remains stalled.
Sentiment
Mostly positive on efficiency gains for agents, with notes on Pro delay and specialized Cyber tier
“The signal in Google's release today is the framing: both new models are explicitly positioned for production AI agents. ... The efficiency race is the frontier. ... Google's 3.6 Flash knowledge cutoff moves to March 2026... My read: the 83% computer use score and the token efficiency improvements point to Google closing in on Anthropic's lead in agentic benchmarks.”
“New from Google DeepMind: Gemini 3.5 Flash Cyber, a version of 3.5 Flash 'fine-tuned to find, validate, and patch vulnerabilities quickly and efficiently.' Notably, it's getting a similar staged release as Mythos and GPT-5.6... 'This will give frontline defenders a head start in finding and fixing critical vulnerabilities before they can be exploited, while mitigating against broader misuse.'”
“【速報翻訳】Google が Gemini 3.6 Flash・3.5 Flash-Lite・Cyber の 3モデル同時公開 ・3.6 Flash: トークン 17% 削減で同コスト（効率化） ・3.5 Flash-Lite: 高スループット・低レイテンシ、$0.30/Mtok（極安） ・Cyber: セキュリティ特化、脆弱性検出・修復 トークン削減×同価格は「モデル枯れ期」の信号。コスト競争が本格化。”
“Google quietly changed the economics of Gemini Flash. ... The biggest catch: according to Artificial Analysis, the cost per task for 3.5 Flash Lite jumped from $0.04 to $0.09 despite the model improvements. Somewhere inside Google, Gemini 3.5 Pro is probably crying. Do you think the quality gains justify these new prices?”
Split
~60/40 positive on efficiency vs. concerns over pricing changes and Pro delay; practitioners focus on agent costs, analysts question value.
Sources
- Google releases three new Gemini models — but no 3.5 ProTechCrunch
- Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash CyberGoogle
- Google Releases 3 New AI ModelsX
- Google Releases Gemini 3.6 Flash and 3.5 Flash CyberX
- Gemini last models: temperature, top_p, and top_k are deprecated and ignoredAi
- Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash CyberGoogle
- Introducing Gemini 3.5 Flash CyberGoogle DeepMind
- Right on Cue, Google Is Reportedly on the Verge of Shipping a Coding-First AI Model - GizmodoGizmodo
- Introducing Gemini 3.8 Flash and 3.8 Flash CyberGoogle
- Google releases Gemini 3.8 Flash, its third Flash model in six weeksArs Technica
- Google says its new Gemini 3.8 Flash model ‘works harder’ but might cost moreThe Verge