If you've asked Gemini — in Google Search, in the Gemini app on your phone, in Gmail's smart reply — about something from last year and gotten an answer that felt strangely incomplete or dated, that probably wasn't the model getting the facts wrong. It was the model not knowing the facts at all.

Every AI model has a knowledge cutoff: a date past which it learned nothing new. Gemini 3.5 Flash, the fast workhorse model powering a huge chunk of Gemini-based apps and features, had its cutoff set at January 2025. Which means for most of the last year and a half, the model answering your questions didn't know about anything that happened after that date. Think of it like a knowledgeable colleague who went completely off the grid in early 2025. Still knows a lot. Just needs significant updating before you'd trust their read on current events.

Google released Gemini 3.6 Flash on July 21. The knowledge cutoff moves to March 2026. Fifteen months of catch-up, in one release. The model also costs less per task than its predecessor — $7.50 per million output tokens, down from $9.00, and it uses 17% fewer tokens to do equivalent work, making the effective savings larger than the sticker suggests. The benchmark improvements are real too. More on all of it below.

Google also released Gemini 3.5 Flash-Lite (a lighter, cheaper variant) and Gemini 3.5 Flash Cyber, a security-tuned model restricted to governments and trusted partners. Flash Cyber is the quietest part of the announcement. And Google teased Gemini 4, with no dates or details attached.

Source spread

• Google DeepMind — official announcement [hype]: Google's framing leads with benchmark gains and new capabilities; the knowledge cutoff update gets a line, not a headline.
• GCN — cybersecurity model coverage [builder]: Independent coverage focused on Flash Cyber and the specific benchmark numbers; most rigorous on the output-token efficiency claim.
• Officechai — pricing comparison [builder]: Price comparison between 3.5 Flash and 3.6 Flash; confirms the output pricing drop.
• Trilogy AI Substack — "Real cost drop is bigger than the sticker" [skeptic]: Independent pricing analysis; works out the combined savings from rate cut plus token efficiency.
• 9to5Google — Gemini 4 tease [hype]: Focused on the forward-looking Gemini 4 mention; less on current-model specifics.

What's real:

• The knowledge cutoff jump is the headline, whether or not Google framed it that way. January 2025 to March 2026 means the model now has working knowledge of most of what happened in AI, product launches, world events, and company news over the last 15 months. For anything where recency matters — and that's a surprising amount of what people ask AI about — this is the most meaningful change.
• The benchmarks moved across multiple domains, not just one. DeepSWE up 12 points, MLE-Bench up 14, OSWorld-Verified up 4.6. When improvements show up across coding tasks, ML research tasks, and computer-use tasks simultaneously, it's harder to argue they're benchmarking artifacts.
• The effective cost cut is bigger than $7.50 vs $9.00. Seventeen percent fewer output tokens on the same prompts means your actual bill running 3.6 Flash is meaningfully lower than the rate card difference suggests. You get the price cut and the efficiency gain.

What deserves a side-eye:

• Flash Cyber's access restrictions aren't explained. "Governments and trusted partners" is a policy position, not a technical description. Something about Flash Cyber isn't safe for general access. The company hasn't said what.
• March 2026 still trails real-time by four-plus months. The knowledge update matters. It doesn't make the model current. News from April onward isn't in the training data.
• The Gemini 4 tease is a tease. No date, no specs, no timeline. It's either a signal that something good is close or a hedge against competitive pressure from GPT-5.6 and Fable 5. I don't know which.

Samwise's take

What to do about it

• If you use Gemini in Google apps (Docs, Gmail, Search): The knowledge cutoff improvement will roll out gradually as Google updates the underlying models in consumer products. The API version is already updated. Don't expect an instant change everywhere, but if current-events awareness has frustrated you, give it another try in a few weeks.
• For anything time-sensitive, still cross-check. The new cutoff is March 2026. That's better than January 2025, but it's not today. For questions where the last four months matter — recent regulations, recent product releases, recent company news — use Google Search as a source, not just the AI answer.
• If you're a developer on the Gemini API: Recalculate your token costs before assuming savings. On output-heavy prompts, the combined effect of the rate cut and 17% efficiency gain can be significant. Test with a sample batch before migrating production traffic.
• On Flash Cyber: If you're building in a security, government, or regulated-industry context, check whether you qualify for access. The capabilities Google is keeping restricted are almost certainly interesting for legitimate security workflows.

Further reading

• Google DeepMind — Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — primary source, all benchmark numbers sourced here
• GCN — Google launches Gemini 3.6 Flash and cybersecurity model — independent coverage with specific benchmark detail
• Officechai — Gemini 3.6 Flash pricing vs 3.5 Flash — pricing comparison
• Trilogy AI Substack — The real cost drop is bigger than the sticker — independent cost analysis
• 9to5Google — Google launches Gemini 3.6 Flash, teases Gemini 4 — consumer framing and Gemini 4 tease