Google ships Gemini 3.8 Flash: matches Opus 5 on several agent benchmarks at ~15% of the price
AGI Hunt · wechat · 2026-09-03
Google DeepMind released Gemini 3.8 Flash and a security-focused Gemini 3.8 Flash Cyber — the third Flash update in six weeks, per Logan Kilpatrick.
Benchmarks: 73.7% on DeepSWE v1.1 (nearly matching Claude Opus 5's 74.0%, up from 65.3%), 89.4% on Terminal-bench 2.1 (edging Opus 5's 89.1%), and field-best scores on Harvey's legal agent benchmark (10.0%) and Vals Finance Agent v2 (61.4%). Weak spots remain: 19.1% on Terminal-bench 4.0 (vs Opus 5's 51.8%) and 59.0% on OSWorld-2.0 (vs 75.4%).
Pricing matches 3.7 Flash: $0.75/1M input, $3.75/1M output (promo until end of 2026) — roughly 15% of Opus 5's $5/$25, suited for large-scale moderate-difficulty agent workloads. The Cyber variant targets frontier vulnerability detection and auto-patching. Available on Gemini API, AI Studio, Antigravity, and the Gemini App.
Related event: Google Launches Gemini 3.8 Flash and Security-Focused Flash Cyber(60 posts)→
More from Models
- ChatGPT helps resolve 30-year-old stable forking conjecture in logic — Dr_Singularity · 2026-09-03
- Users praise Gemini 3.8 Flash: "coding is fun again" — xennygrimmato_ · 2026-09-03
- Grok Bot's Token Limits Frustrate Users; Subscriptions With All-You-Can-Eat Tiers May Win — NickPassig · 2026-09-03
- Author Burkov says Anthropic ignored his refund requests, so he cobbles together Codex and Grok — burkov · 2026-09-03
- Google launches Gemini 3.8 Flash Cyber security model alongside Fairwind Program for defenders — GoogleAI · 2026-09-03
- Matt Shumer: slow AI releases aren't a wall — safety clearance is the bottleneck, wave of frontier models imminent — mattshumer_ · 2026-09-03