On September 2, 2026, Beijing time, Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. Tulsee Doshi, senior director of product management, and Raluca Ada Popa, head of Gemini security at Google DeepMind, co-signed the announcement. The Standard Edition is geared toward software engineering, agent tasks, and professional knowledge workflows; the Cyber Edition is geared toward vulnerability discovery and automatic patching, and only provides priority access to qualified defenders through the newly established Fairwind Program. The two are not the same service for the same user group.
Technical Path and Price Boundary
Gemini 3.8 Flash is based on the 3.7 Flash released three weeks ago. Google said that the two 3.8 variants share basic capabilities and are repeatedly evaluated and adjusted by long-running agent loops; the standard version will add reasoning steps and iteratively call tools when facing complex tasks. Developers can still adjust the effort level and choose between quality, cost and latency; if computing efficiency is the primary constraint, Google recommends lowering the effort level or continuing to use 3.7 Flash, which is still supported. The announcement did not disclose parameter quantities, a list of training data, or more specific architectural changes.
The entry API price for the Standard Edition as of December 31, 2026 is US$0.75 per 1 million input tokens and US$3.75 per 1 million output tokens, which is the same as the entry unit price of 3.7 Flash. Google's footnote states that the official prices will rise to US$1.50 and US$7.50 respectively from January 1, 2027. The same unit price does not mean that the total cost of each task is the same: the official explanation is that 3.8 may use more tokens in complex tasks, especially at higher effort levels, but does not provide the average usage, cost range of various tasks, or comparison with the bill for the same task in 3.7.
Model specifications and distribution channels
The Google DeepMind model card states that Gemini 3.8 Flash can receive text, images, audio and video, the maximum context window is 1 million tokens, the output is text and the upper limit is 64,000 tokens. The deadline for model knowledge is marked as March 2026, but the card also states that information in some areas may be updated by this time, and in other areas it may only be January 2025. The model card describes 3.8 as a subsequent version of 3.7, and points the architecture, training data processing, hardware and software instructions to the 3.7 model card. No independent parameter scale or training data details are announced in this material.
Developers can use the standard version in Gemini API, Google AI Studio, Android Studio and Google Antigravity, and enterprises can access it through Gemini Enterprise. Google also said that Google AI Pro and Ultra subscribers can use 3.8 Flash in Gemini applications, and consumer portals also include Google Search’s AI Mode and Google Sheets. The official release does not list country-by-country availability tables, free quotas, call limits for each portal, or phased coverage times. Therefore, the existing information can only confirm these distribution channels, and it cannot be inferred that all regions, account types, and product portals have the same permissions.
Cyber Edition Admission and Assessment
Gemini 3.8 Flash Cyber focuses on vulnerability discovery and automatic patching. Google said it prioritizes investing in vulnerability fixes rather than offensive capabilities such as exploiting vulnerabilities; this version adopts looser mitigation measures for network security tasks than the standard version, and therefore does not enter public channels for all users. The Fairwind Program accepts applications from trusted government agencies, critical infrastructure operators and software maintainers and provides priority access to accredited defenders. The announcement did not announce review standards, participation quotas, service prices, specific regions or a timetable for opening to general developers.
In the evaluation published by Google, the standard version obtained 54.9% in HLE-Verified; the Cyber version achieved a success rate of over 70% in the internal vulnerability discovery benchmark covering 20 programming languages, and the pass@1 was 47.2% in the CWE-Bench patching benchmark run by Collinear. Google also said that the number of correct vulnerability patches generated by the Chrome security team using the Cyber version reached 2.6 times that of the best large-scale commercial model in the comparison object, and the Google Cloud vulnerability research team found a key basic vulnerability in less than 2 hours. Except for the operator information of CWE-Bench, these numbers and cases did not provide complete external reproduction with the announcement.
Security restrictions and subsequent nodes
The model card lists hallucinations, occasional slowdowns or timeouts, and the possibility of higher effort levels costing more tokens as known limitations of Standard. The card explains that the security results listed in the development stage come from automated assessments, not manual assessments or red team testing; the multi-language security indicators are 5.4 percentage points lower than those in 3.7 Flash, and the lower the indicator, the better. This change only corresponds to specific automated indicators. Google said that manual review found that most of the flagged losses were false positives or non-serious problems; the model card also said that the standard version has no substantial new cutting-edge security framework capabilities compared to 3.7. This conclusion does not apply to the restricted Cyber version.
The Verge confirmed in a public information stream on September 2 that Gemini 3.8 Flash has been released, and based on Google's instructions, it summarized its efforts to add reasoning steps and iteratively call tools in complex tasks. The media summary also distinguishes between the introductory unit price and the actual cost: although the unit price is the same as 3.7, more tokens may still drive up the total cost of a mission. There are two clear nodes to check: the entry price of the standard version ends on December 31, and the official price will take effect on January 1, 2027; the Cyber version continues to accept applications from specific defense agencies through the Fairwind Program, and Google has not yet announced an opening date for general users.