Gemini 3.8 Flash

What it is

Gemini 3.8 Flash is a model in Google's Flash line, released on 2 September 2026. It was published together with a cybersecurity variant, Gemini 3.8 Flash Cyber. The announcement post was written by Tulsee Doshi and Gemini Security Lead Raluca Ada Popa.^[1] Flash is Google's family of models built to answer quickly and at low cost rather than to be the largest model on offer, so 3.8 Flash is an upgrade inside that family rather than a new tier of its own.

Positioning at launch

At launch Google called it its best reasoning and coding model so far, offered at the speed and low cost of 3.7 Flash.^[1] The company also noted that it was the third Flash release in six weeks, which says something about how fast this line was being revised.^[1] On HLE-Verified, a benchmark built from hard questions, the model scored 54.9%.^[1] That is the benchmark figure the launch relied on when making the reasoning claim.

Pricing

Introductory pricing runs through 31 December 2026 at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens. From 1 January 2027 those rates rise to $1.50 and $7.50.^[1][3] Two further tiers were published for the same window. Batch and Flex pricing was $0.375 per 1 million input tokens and $1.875 per 1 million output tokens through 31 December 2026, rising to $0.75 and $3.75 from 1 January 2027.^[3] Priority pricing was $1.35 per 1 million input tokens and $6.75 per 1 million output tokens through 31 December 2026, rising to $2.70 and $13.50 from 1 January 2027.^[3] In every tier the introductory rate is exactly half the later rate, so the published schedule doubles the unit price at the start of 2027. Tokens are the units of text a model reads and writes, and metering them is how API access is billed.

Where it is available

At launch the model could be reached in Antigravity, the Gemini API, Android Studio, Stitch and Gemini Enterprise.^[1] It was also made available to Google AI Pro and Ultra subscribers in the Gemini app, in AI Mode, and in Gemini in Google Sheets.^[1] That spread covers developer tooling, an agent-oriented environment, enterprise access and consumer surfaces at the same time, which is typical of how Google ships a Flash release.

The Cyber variant

Gemini 3.8 Flash Cyber is not a general release. It was limited to trusted defenders through the Fairwind Program.^[1] On CWE-Bench patching it scored 47.2% pass@1, and on an internal vulnerability-discovery benchmark it scored over 70% across 20 languages.^[1] Google reported the CWE-Bench figure next to a comparison: 47.2% for its model against 47.8% for a leading frontier model.^[1] Pass@1 means the share of tasks solved on the first attempt, so the number describes how often a single try produced a working patch. Restricting a security-focused model to vetted defenders is a way of keeping a capable tool out of the hands of the people it is meant to be used against.

Follow-on models and routing

More models arrived in the API later that month. On 15 September 2026 came gemini-3.8-live and gemini-3.8-live-extended-thinking, and on 22 September 2026 came gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts.^[2] The names mark out live and speech-output variants built on the same generation. Google then designated gemini-3.8-flash as the auto-routing target for gemini-2.5-pro requests from 1 December 2026.^[2] Auto-routing means a request that does not name a model is sent to a chosen default; here, traffic previously aimed at an older model would land on 3.8 Flash instead.

What the release shows

The release is a snapshot of how a model family is kept current. A launch post sets the claim, a security-specific sibling shows the same base being tuned for one narrow job under access controls, a price sheet sets a discount window with a known expiry, and later API entries fill in live, speech and lite versions. The routing change is the part that touches ordinary users most quietly: when an older model's requests are pointed at a newer one, people who never chose a model start getting different answers.