Sunday, 27 September 2026
Anime, manga, and games, with a take · A Yukimedia publication
← all stories other 3 sources · Jul 22 · · Updated

Google Announces Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Google is releasing production-oriented models optimized for agent workloads rather than a flagship upgrade, signaling a shift toward practical efficiency gains as AI spending faces increased scrutiny.

Key Facts

  • Gemini 3.6 Flash reduces output tokens by 17% compared to Gemini 3.5 Flash on the Artificial Analysis Index.
  • Gemini 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens.
  • Gemini 3.5 Flash-Lite generates 350 output tokens per second and is priced at $0.30 per million input tokens and $2.50 per million output tokens.
  • Gemini 3.5 Flash Cyber found 55 unique confirmed issues in the V8 JavaScript engine, compared to 47 for Gemini 3.5 Flash and 36 for Claude Opus 4.6.
  • Google confirmed that pre-training for Gemini 4 has begun, described as its most ambitious pre-training run to date.

Reporting from 3 sources: ASCII.jp, GameBusiness.jp, GIGAZINE.

Google Announces Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Google announced three new models in its Gemini Flash series on July 21, 2026: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Gemini 3.6 Flash is a general-purpose model that reduces output tokens by 17% compared to the previous Gemini 3.5 Flash, while improving coding and multimodal performance. It scored 49% on the DeepSWE coding benchmark, up from 37%. Pricing is $1.50 per million input tokens and $7.50 per million output tokens, with output cost reduced from $9. Gemini 3.5 Flash-Lite is the fastest model in the 3.5 series, generating 350 output tokens per second, and is priced at $0.30 per million input tokens and $2.50 per million output tokens. Gemini 3.5 Flash Cyber is a cybersecurity model built on Gemini 3.5 Flash, fine-tuned to find and fix software vulnerabilities. It is integrated into Google's CodeMender agent and will initially be offered as a limited trial to government agencies and trusted partners. Google also confirmed that Gemini 3.5 Pro is still in testing with partners, and pre-training for the next-generation Gemini 4 has begun. The announcement received mixed reactions, with some praising speed and efficiency while others questioned cost performance versus competitors.

Google is developing an internal AI chip called "Frozen v2," which could be 6 to 10 times more efficient than existing chips when comparing tokens generated per unit of power, according to a report from The Information. TechCrunch, which asked Google about the report, received a comment: "Our teams are always researching and experimenting with new innovations to deliver the best performance and efficiency to our users and customers. Not every project becomes a product, but this thorough exploration is core to our full-stack approach." The chip is expected to be released sometime in 2028.

Reactions to the model announcements were mixed. On X, some users highlighted that Gemini 3.6 Flash "reduces token usage while improving performance in complex agent processing and multimodal tasks." A developer using it for an app that estimates meal calories from images praised its image understanding. But on the independent evaluation site Artificial Analysis, Gemini 3.6 Flash scored 50 on the "Intelligence Index," and the community noted that other models achieve equal or higher scores at lower cost. Critical posts on X said it is "cheap but lags behind competitors in major coding evaluations."

Engineer Simon Willison tested Gemini 3.6 Flash on a benchmark called "Pelican on a Bicycle," asking it to generate an SVG image. The model produced a bicycle with a visible shape and pedaling motion, but the pelican's beak looked strange and its bottom appeared fused with the saddle. Willison noted that older models can sometimes output a better pelican on a bicycle, so the test cannot be compared simply. AI app builder Playcode ran the same test and found the output was "only slightly different" from Claude Fable 5 at about one-fifth the cost.

Synthesized by Yomimono from the 3 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.

More on this

Sources