Google Quietly Dropped Three Gemini Models Yesterday and Started Training Gemini 4. The Flagship Delay Story Just Got Worse.

Hey guys, Mr. Technology here.

It is Wednesday, July 22, 2026. The headlines are still chewing on the Gemini 3.5 Pro delay. What Google actually did yesterday was ship three new models at once and announce pre-training has begun for Gemini 4. That is a strategy pivot story, not a delay story.

The three: Gemini 3.6 Flash (workhorse, GA today), Gemini 3.5 Flash-Lite (cheapest, fastest 3.5-class model Google has shipped), and Gemini 3.5 Flash Cyber (cybersecurity specialist inside CodeMender, gated to governments and "trusted partners"). Read them together: Google has stopped treating Flash as the cheap tier and started treating Flash as the shipping tier. Five days ago I wrote 3.5 Pro "does not exist on a timeline that does not exist." Yesterday confirmed the trajectory.

What Actually Shipped

Gemini 3.6 Flash — GA today, cheaper than 3.5 Flash. 17% fewer output tokens on Artificial Analysis, up to 65% on DeepSWE. $1.50 input / $7.50 output per 1M vs 3.5 Flash's $1.50 / $9.00. DeepSWE 49% vs 37%, OSWorld-Verified 83.0% vs 78.4%, GDPval-AA v2 1421 vs 1349.

Gemini 3.5 Flash-Lite — the cheapest frontier-class API on the market. 350 output tokens/sec at $0.30 / $2.50 per 1M. Terminal-Bench 2.1 54% vs 3.1 Flash-Lite 31%. SWE-Bench Pro 54.2% beats 3 Flash at 49.6%. OSWorld-Verified 74.0% crushes 3 Flash at 65.1%. This is a frontier model at commodity price — not "lite" in the way the industry has used that word. Grok 4.5 and Kimi K3's $0.30 cached pricing opened the door. Flash-Lite just won.

Gemini 3.5 Flash Cyber — gated specialist inside CodeMender. Fine-tune of 3.5 Flash for vulnerability discovery. Not a public API — exclusive to governments and "trusted partners" via CodeMender (DeepMind's security agent, October 2025). On Google's Big Sleep eval against Chrome and Safari it "significantly surpassed" 3.5 Flash, 3.6 Flash, and Claude Opus 4.6. On Chrome production commit scanning: 55 unique V8 engine issues versus 47 for 3.5 Flash and 36 for Opus 4.6, including 10 neither other model caught. Google Cloud Vulnerability Research found RCE bugs in public APIs in under two hours.

Why The Gating Matters

Dual-use is the 2026 reality. Anthropic hit this when Opus 4.6 started refusing cyber-offense evals — which is also why Google omitted Opus 4.6 from the V8 chart. OpenAI hit it with GPT-5.5 Daybreak. Google picked a third path: ship the model, do not ship the API.

CodeMender is the wedge product. Not a chat surface — an autonomous security agent. Tying 3.5 Flash Cyber to CodeMender turns a model release into an enterprise security product. Wiz, Cloud CISO, and Google's own Chrome/Android/Cloud/Ads/YouTube teams are already using it.

It sets the precedent. Expect 3.5 Flash Med, 3.5 Flash Legal, 3.5 Flash Finance, 3.5 Flash Gov. The Flash family is the new platform; the specialists are the products you cannot buy.

Hidden Story: Gemini 4 Is In Pre-Training

Buried in yesterday's blog: "We have started our most ambitious pre-training run yet, for Gemini 4." That reframes the 3.5 Pro delay. The goalposts have moved to Gemini 4, and every Flash-tier release in between is filler to keep customers inside Google Cloud while the real flagship trains. The table tells the rest.

Google now ships the cheapest frontier-class API in the industry, a second Flash variant as a price-war counter-punch, and a gated security specialist nobody else has. They are also the only major lab without a Pro-tier flagship and the only one with a confirmed 4-class successor in training.

What You Should Actually Do

One — migrate tier-one Google workloads from 3.5 Flash onto 3.6 Flash today. Same input price, lower output price, 17-65% fewer output tokens, better benchmarks.

Two — re-evaluate every "lite" workload against Flash-Lite. At $0.30 / $2.50 you can route summarization, classification, retrieval rewrite, and high-volume tool-calling for less than GPT-4o-mini cost in 2024.

Three — stop waiting on Gemini 3.5 Pro. It is a parallel-track effort, not the next model. Treat GA as opportunistic. Re-anchor Google workloads on Flash and use Fable 5 / Sonnet 5, GPT-5.6 Sol, Kimi K3, or Inkling for default route.

Four — if you do cybersecurity AI, get into the CodeMender pilot or build against Claude Mythos / GPT-5.5 Daybreak today. Talk to your Google Cloud account team now.

Five — price vendor risk against Gemini 4, not 3.5 Pro. Q1 2027 is the realistic Gemini 4 window. Hedge with a second vendor.

Where I'd Pump the Brakes

V8 number is impressive but not independently audited. Google wrote the eval, ran it, chose what to publish. Wiz and Cloud CISO feedback is quoted, not reproduced.

Gating can become lock-in. Read the data-residency terms before you sign anything.

Cheap is not cheap-forever. Flash-Lite is a price-war move. Expect it to slide lower if Anthropic or OpenAI squeeze.

Mr. Tech's Take

The 3.5 Pro delay was never the real story. The real story is that Google shipped a complete Flash-tier stack and started training Gemini 4 instead of fixing the Pro-tier problem. That is prioritization under pressure, not failure.

Anthropic and OpenAI just lost the cheapest-API crown. Sonnet 5 and GPT-5.6 Sol were the June leaders. They are not anymore. Flash-Lite sets the new floor. Both labs will drop mini-class prices within weeks, or absorb low-end volume erosion.

Specialist gated models are the new product surface. OpenAI has GPT-5.5 Daybreak. Anthropic has Claude Mythos. Google has 3.5 Flash Cyber. The public API is becoming the commodity tier. The product is becoming the gated specialist. Your 2027 procurement has to account for this — you will not be buying raw models, you will be buying agent products with model identity hidden behind them.

Gemini 4 is in pre-training. That puts Google on the same Q1 2027 horizon as Claude Fable-class successors and GPT-6 family rumors. The frontier model cycle is collapsing from 18 months to ~6. The labs that win 2027 will ship a flywheel of cheap commodity API + gated specialist agents + flagship refreshes on a six-month clock.

The headline is that Google has now built the flywheel. As of 5pm PT on July 21, 2026, only one lab had all three layers live. That is the real announcement.

— Mr. Technology

Release date: July 21, 2026 (Tuesday)

Models released:

• Gemini 3.6 Flash — GA in Gemini API, Gemini Enterprise, AI Studio, Android Studio, Antigravity, Gemini app. $1.50 / $7.50 per 1M. 17% fewer output tokens than 3.5 Flash; up to 65% on DeepSWE. DeepSWE 49% vs 37%; MLE-Bench 63.9% vs 49.7%; OSWorld-Verified 83.0% vs 78.4%; GDPval-AA v2 1421 vs 1349.
• Gemini 3.5 Flash-Lite — GA today. 350 output tokens/sec. $0.30 / $2.50 per 1M. Terminal-Bench 2.1 54% vs 31%; SWE-Bench Pro 54.2% beats 3 Flash 49.6%; OSWorld-Verified 74.0% beats 3 Flash 65.1%.
• Gemini 3.5 Flash Cyber — 3.5 Flash fine-tune for cybersecurity. Lives inside CodeMender. Governments and trusted partners only via limited pilot. 55 unique V8 issues vs 47 / 36 for 3.5 Flash and Opus 4.6.

Additional: Pre-training for Gemini 4 has commenced. Gemini 3.5 Pro remains in partner testing with no public date.

Other frontier models for context: Claude Fable 5 / Sonnet 5 ($5 / $25), GPT-5.6 Sol ($5 / $30), Kimi K3 ($3 / $15, open-weights July 27), Inkling (Apache 2.0 self-host), Grok 4.5.

Sources: Google Blog — 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber · DeepMind — Introducing Gemini 3.5 Flash Cyber · The Hacker News coverage · Reuters — Google updates lightweight Gemini, flagship delayed · DataCamp — Gemini 3.6 Flash benchmarks · OpenRouter — Gemini 3.6 Flash · DeepMind — 3.6 Flash model card · DeepMind — 3.5 Flash-Lite model card · llm-stats.com updates feed.