What Google's Own Launch Post Actually Says
Published: September 9, 2026 | Category: AI | By Mahesh | Source: Google's official launch post, primary data current as of September 9, 2026
Depth Grid is starting a new regular feature: a genuinely new AI model, app, agent or piece of production technology, read straight from the company's own announcement rather than the usual churn of secondhand summaries. First up is Google's Gemini 3.8 Flash, which the company's own blog post, published September 2, 2026 and authored by Senior Director of Product Management Tulsee Doshi and Gemini Security Lead Raluca Ada Popa, describes as Google's "best reasoning and coding model yet, at the same speed and low cost of 3.7."[1] It is also, by Google's own account, its third Flash-tier release in six weeks, following Gemini 3.7 Flash on August 13, 2026.[1]
What Gemini 3.8 Flash Actually Is
Claim: Gemini 3.8 ships as two distinct models built on the same underlying intelligence, aimed at different users. Source: Google's own launch post states the release "introduces 2 variants," describing Gemini 3.8 Flash as "our most intelligent workhorse model, delivering significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains," and Gemini 3.8 Flash Cyber as "our most capable cybersecurity model with frontier-level performance in vulnerability detection and automated patching."[1] The post further states that "while tailored for different deployment environments, both of today's releases are powered by the same foundational intelligence, and further accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models."[1] Analysis: The decision to ship a general model and a security-hardened variant from the same base, rather than one model for everyone, is itself the more interesting product signal here. It suggests Google now treats cybersecurity capability as something that needs its own access controls and permissions layer, not just a feature toggle, a pattern that shows up elsewhere in frontier AI releases this year. Published: September 2, 2026.
The Benchmark Claims, Read From Google's Own Post
Claim: Google reports 3.8 Flash outperforming 3.7 Flash and approaching costlier frontier models on long-horizon coding and specialised reasoning tasks. Source: The launch post states that "on DeepSWE v1.1 (Long-Horizon Software Engineering), 3.8 Flash outperforms most larger frontier models in autonomously solving complex engineering problems end to end, only at a fraction of the cost," and that the model "achieves a 54.9% on HLE-Verified, demonstrating its ability to handle multi-step reasoning across STEM, humanities, and professional fields."[1] Google also reports 3.8 Flash outperforming 3.7 Flash and other frontier models on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, two third-party benchmarks it links directly in its own post rather than citing without attribution.[1] Analysis: Linking to independently run, third-party benchmark leaderboards rather than only citing internal test results is a genuinely useful transparency signal, since it lets outside developers verify the same numbers Google is quoting rather than taking the company's word for it. That said, these remain benchmark results, not a substitute for testing the model against your own actual workload, and Google's own post is explicit that the gains come from the model "working harder," not a new base architecture, a distinction worth remembering before assuming the benchmark gains translate directly to every use case. Published: September 2, 2026.
The Cyber Variant Almost Nobody Can Access
Claim: Gemini 3.8 Flash Cyber shows strong results on real-world vulnerability detection but is deliberately restricted to a small set of vetted organisations rather than being generally available. Source: Google's post states the model "showcases an impressive leap over our previous models and reaches a success rate exceeding 70%" on an internal benchmark spanning vulnerability discovery across 20 programming languages, and separately reports that "the Chrome Security team found that 3.8 Flash Cyber produced 2.6 times more correct patches to vulnerabilities in Chrome than the best commercial models that are much larger," while security firm Wiz "found that Gemini 3.8 Flash Cyber achieves +7.5-9.7% higher recall on their internal penetration testing benchmark for a 2.3-5.2x lower cost."[1] Access is restricted to the Fairwind Program, which Google describes as providing "trusted government authorities, as well as critical infrastructure operators and software maintainers with prioritized access."[1] Analysis: Restricting the most capable offensive-adjacent AI security tooling to vetted defenders rather than releasing it broadly is a deliberate, disclosed policy choice, not an oversight, and it reflects a genuine industry-wide tension this year between making powerful defensive AI tools available fast enough to matter and limiting how easily the same capability could be repurposed by an attacker. Anyone running critical infrastructure or a large codebase can apply for access directly through the program link Google provides. Published: September 2, 2026.
The One Honest Caveat Google Includes Itself
Claim: Google explicitly tells some developers not to upgrade. Source: The launch post states plainly: "for applications where compute efficiency is the primary constraint, developers can utilize lower effort levels to minimize token overhead or continue to rely on Gemini 3.7 Flash, which remains fully supported for efficiency-first workloads."[1] Analysis: A vendor telling developers their older, cheaper model remains a legitimate choice is unusual enough to be worth flagging on its own. It is a direct acknowledgement that the performance gains come from the model doing more work per task, meaning real-world cost and latency can rise for workloads where the extra reasoning steps do not translate to a better outcome. Teams evaluating whether to switch should benchmark token usage and total cost per task on their own workload before assuming the new model is automatically the better economic choice, not just the more capable one. Published: September 2, 2026.
What It Actually Costs to Use in India
Google's own post states that in the consumer Gemini app, 3.8 Flash "is available to Google AI Pro and Ultra subscribers."[1] In India, that means the Google AI Pro plan at ₹1,950 a month, or the AI Ultra plan starting at ₹6,500 a month, both billed natively in rupees with UPI accepted, according to pricing trackers citing Google's official Indian subscription pages.[2][3] Developers building directly against the API are not required to hold a subscription at all: the $0.75 per million input token and $3.75 per million output token introductory rate applies through Google AI Studio and the Gemini API regardless of geography, and Google's free tier for Gemini 3 Flash and Gemini 3.1 Flash-Lite remains accessible through AI Studio with no credit card required, which pricing guides describe as an unusually generous free API tier compared with OpenAI or Anthropic, both of which charge from the first token.[4]
How to Actually Try It
Google's own post lists the access paths directly, by user type, rather than leaving it to third parties to explain:
- Developers: build with 3.8 Flash in Google Antigravity, or start directly in the Gemini API through Google AI Studio and Android Studio, or generate interfaces in Stitch. Full setup instructions are in Google's own developer docs.[1]
- Enterprises: access 3.8 Flash through Gemini Enterprise.[1]
- Consumers: available to Google AI Pro and Ultra subscribers in the Gemini app, AI Mode in Google Search and Gemini in Google Sheets.[1]
- Security teams: apply for prioritised access to the Cyber variant through the Fairwind Program application page.[1]
Common Questions
Sources
- Tulsee Doshi and Raluca Ada Popa, "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber," Google, official launch post, September 2, 2026. blog.google
- Technosports, "Google AI Ultra Is Now Available in India at ₹6,500/Month," citing Google's official India subscription pricing, May 2026. technosports.co.in
- Equity Research India, "Google Gemini AI Pricing in India 2026," citing Google's official Indian INR pricing, August 2026. equityresearchindia.com
- Equity Research India, "Google Gemini AI Pricing in India 2026," free API tier comparison (see source 3).
Read More on Depth Grid
- OpenAI's New Model Design Makes It Harder to Read Its Mind
- Why the AI Moat Is Shifting to the Workflow
- What Big Tech's Own Filings Say About the AI Spending Bet
- The Compute and Power Bottleneck: AI's Real Limit
- Enterprise AI Agents: Where the Big Checks Go
Article by Mahesh | Depth Grid