Full Breakdown
Google Unveils Three New Gemini AI Models, Delays Flagship Pro Update
7/22/2026, 3:46:03 AM
New Gemini Models Launched on July 21
On July 21, Google DeepMind announced three Gemini models built on the Gemini 3.5 Flash foundation: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. All are accessible through the Gemini API, AI Studio, Enterprise, and the Gemini app, except Flash Cyber, which is limited to governments and vetted partners in a pilot.
Background: Competitive Pressure and Pro Model Delay
Google’s releases come amid upgrades from rivals OpenAI and Anthropic, which have introduced GPT-5.5, GPT-5.6, Claude Opus 4.8, and the Mythos security line. Google had signaled a June launch for its flagship Gemini 3.5 Pro during its May I/O keynote, but internal testing fell short, leading to an undefined delay. Product lead Logan Kilpatrick confirmed that Gemini 3.5 Pro is still being tested with partners while the team begins pre-training for Gemini 4.
Model Details and Performance Metrics
| Model | Primary Focus | Token-efficiency | Pricing (per million tokens) | Speed / Benchmarks |
|---|---|---|---|---|
| Gemini 3.6 Flash | General-purpose coding, knowledge work, multimodal tasks | 17 % fewer output tokens vs. Gemini 3.5 Flash; up to 65 % reduction on DeepSWE benchmark | $1.50 input / $7.50 output | Improves coding and document-analysis scores; lower latency |
| Gemini 3.5 Flash-Lite | High-volume, low-latency workloads | Fastest in the 3.5 series, 350 output tokens / second | $0.30 input / $2.50 output | Beats Gemini 3.1 Flash-Lite on Terminal-Bench 2.1 (54 % vs. 31 %) and SWE-Bench Pro (54.2 % vs. 49.6 %) |
| Gemini 3.5 Flash Cyber | Automated detection and remediation of software vulnerabilities | Lower per-token cost than larger models; competitive on CyberGym benchmark | Not publicly priced; limited-access pilot | Integrated with Google’s CodeMender agent for security reports |
Google notes that token-efficiency gains translate into cost savings for large-scale workloads, and Flash-Lite’s speed suits chatbot infrastructure and content moderation.
Official Statements from Google
- A Google Cloud spokesperson highlighted the “full-stack approach,” saying co-designing hardware and software enables “highly optimized” performance.
- Logan Kilpatrick said Flash Cyber will initially be offered only to governments and trusted partners because of its dual-use nature, and that the team is “landing soon” with Gemini 3.5 Pro while scaling pre-training for Gemini 4.
Market and Strategic Implications
The announcement arrived just before Alphabet’s quarterly earnings, showing progress despite the Pro delay. Alphabet’s shares slipped 0.66 % to $349.66 afterward. Analysts view the cost-focused lineup as a bid to attract developers and enterprises that prioritize operational economics over frontier reasoning performance, a segment where OpenAI and Anthropic currently lead.
Introducing a dedicated cybersecurity model aims to narrow the gap with Anthropic’s Mythos line, recognized for code-defense capabilities. The limited rollout reflects regulatory concerns around AI-driven vulnerability discovery.
Future Plans and Unresolved Issues
- Gemini 3.5 Pro remains in partner testing with no release date.
- Gemini 4 pre-training has begun; launch timing is undisclosed.
- Access to Flash Cyber is restricted; broader availability depends on pilot outcomes and policy.
- Google is developing a custom AI chip projected to improve Gemini efficiency by up to tenfold, though such hardware typically requires years to mature.
The three new models expand Google’s AI portfolio for production workloads, offering developers a spectrum from ultra-low-cost, high-throughput inference (Flash-Lite) to a mid-range workhorse (Flash 3.6) and a niche security solution (Flash Cyber). Delivering a competitive Pro-tier model soon will be key to maintaining its standing in the generative-AI market.
