Releases·Model Releasefeatured
Gemini 3.8 Flash debuts at 305 output tokens per second, nearly double runner-up Muse Spark 1.2
Google announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on 2026-09-02, its third Flash-tier release in six weeks, positioning 3.8 Flash as its best reasoning and coding model yet at 3.7-level speed and cost. The Cyber variant targets security teams with vulnerability detection and autonomous patching. The release continues Google's cadence of Flash updates while no frontier Gemini Pro has shipped since early 2026.
5434 engagement9 sourcesWed, Sep 2, 2026, 05:41 PM
Developments
Media coverage highlights Google's claim the model 'works harder' but might cost more; users flag 305 tok/s speed charts.
Wed, Sep 2, 2026, 08:11 PM
AI Overview
- Google lists Gemini 3.8 Flash at 305 output tokens per second versus 154 for Muse Spark 1.2 and 126 for GPT-5.6 Luna
- Gemini 3.8 Flash posts an intelligence score of 59, near models scoring 60-66
- Google releases Gemini 3.8 Flash Cyber, tuned for vulnerability detection and autonomous patching, to trusted defenders via its new Fairwind Program
- Gemini 3.8 Flash is Google's third Flash release in six weeks, with no Gemini Pro model since early 2026
- Introductory pricing runs $0.75/$3.75 per million input/output tokens through year-end, rising to $1.50/$7.50
- Google says Gemini 3.8 Flash "works harder" via more reasoning steps and iterative tool calling, but may cost more by using more tokens at higher effort levels
- Google says 3.8 Flash delivers significant gains over 3.7 Flash in software engineering, agentic tasks, and multi-step reasoning, scoring 54.9% on HLE-Verified
- Google reports 3.8 Flash Cyber patches at 47.2% pass@1 on CWE-Bench versus 47.8% for a leading frontier model, and produced 2.6x more correct Chrome patches than larger commercial models
Entities
Related
Social·Community
Google lost its AI lead in 7 months: no Gemini 3.5 Pro, talent exodus, and TPUs handed to rivals
Reddit threads debate why Google lost AI momentum despite unmatched resources, citing no new Gemini Pro model since Gemini 3.1 Pro topped the Artificial Analysis index in February 2026, executive departures, and TPU capacity routed to rivals. A follow-up post reports the Gemini 4 Argon announcement excluded paid Pro-tier users, prompting subscription cancellations.
Reddit threads debate why Google lost AI momentum despite unmatched resources, citing no new Gemini Pro model since Gemini 3.1 Pro topped the Artificial Analysis index in February 2026, executive departures, and TPU capacity routed to rivals. A follow-up post reports the Gemini 4 Argon announcement excluded paid Pro-tier users, prompting subscription cancellations.
Updated Thu, Oct 1, 2026, 03:41 PM — Gemini 4 Argon announced with no access for Pro-tier users, limiting it to Fairwind partners, paid API, and a new Ultra
001391 engagement3 sourcesredditSun, Sep 27, 2026, 04:57 PM
Social·Community
Reddit user breaks down Gemini Flash 3.5-3.8 gains: cost, speed, agentic benchmarks, but worse token efficiency
A Reddit user analyzed Artificial Comparison benchmarks across four recent Gemini Flash releases (3.5 through 3.8) and argues Google has focused each release on a single improvement axis. Comparing Gemini 3.5 Flash to 3.6 Flash, they identify token efficiency and cost reduction as the main change, with 3.6 being the cheapest and 3.5 the most expensive of the four models.
A Reddit user analyzed Artificial Comparison benchmarks across four recent Gemini Flash releases (3.5 through 3.8) and argues Google has focused each release on a single improvement axis. Comparing Gemini 3.5 Flash to 3.6 Flash, they identify token efficiency and cost reduction as the main change, with 3.6 being the cheapest and 3.5 the most expensive of the four models.
2 engagement1 sourceredditThu, Sep 3, 2026, 01:10 PM
Social·Community
Gemini 3.5 Flash-Lite suffices for most users, says developer of three-model deployment
A practitioner running a multi-model AI service for a large institution reports that most users choose Gemini 3.5 Flash-Lite over Gemini 3.8 Flash and Claude Sonnet 5 and express full satisfaction. The post argues entry-level models now exceed former frontier models like GPT-4, making larger models a niche necessity rather than the default for real-world workflows including legal work.
A practitioner running a multi-model AI service for a large institution reports that most users choose Gemini 3.5 Flash-Lite over Gemini 3.8 Flash and Claude Sonnet 5 and express full satisfaction. The post argues entry-level models now exceed former frontier models like GPT-4, making larger models a niche necessity rather than the default for real-world workflows including legal work.
16 engagement1 sourceredditSat, Oct 3, 2026, 04:21 PM
Social·Community
User retrospective compares Gemini 3.5, 3.6, and 3.7 Flash ahead of expected 3.8 Flash
A heavy Gemini user posted a hands-on retrospective of the last three Flash generations on Reddit, anticipating Gemini 3.8 Flash. The comparison finds 3.5 Flash strongest at web search and grounding, while 3.7 Flash leads on writing and creative tasks. The thread offers practitioners subjective but daily-tested impressions of how each generation shifted capabilities.
A heavy Gemini user posted a hands-on retrospective of the last three Flash generations on Reddit, anticipating Gemini 3.8 Flash. The comparison finds 3.5 Flash strongest at web search and grounding, while 3.7 Flash leads on writing and creative tasks. The thread offers practitioners subjective but daily-tested impressions of how each generation shifted capabilities.
51 engagement1 sourceredditMon, Aug 31, 2026, 06:16 PM