A launch event with no Pro. On July 21, Google DeepMind dropped three new models in one stroke — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. No warm-up, no keynote; they simply appeared in Google AI Studio and the Gemini model picker.
Glance at the comment sections, though, and the mood is off. “Worst models I’ve ever seen.” “A local 4B model outperforms it.” “The Kimi K3 team runs on one-150th of Google’s budget and delivered better results.”
So which is it — an efficiency revolution, or a forced draft submission?
Three Models, Three Directions
Gemini 3.6 Flash — the workhorse. Coding and multimodal capability improve, but the headline is token efficiency: identical tasks finish with 17% to 65% fewer output tokens than the predecessor. DeepSWE rises to 49% (from 3.5 Flash’s 37%) and MLE-Bench from 49.7% to 63.9%. Pricing: $1.50 per million input tokens, $7.50 output — slightly under 3.5 Flash’s $9.
Gemini 3.5 Flash-Lite — the speed extreme. 350 output tokens per second, double 3.1 Flash-Lite. Pricing as low as $0.30 input / $2.50 output per million tokens. Yet on SWE-Bench Pro it scores 54.2% — above the standard 3 Flash’s 49.6%.
Gemini 3.5 Flash Cyber — the vertical specialist. Fine-tuned for security vulnerability detection and integrated into CodeMender, available only to governments and trusted partners. Internal testing surfaced 55 issues in the V8 engine codebase, 10 of which other models had missed.
Three models draw a clear web: an efficiency-optimized tier (Flash), a speed extreme (Flash-Lite), a vertical specialist (Flash Cyber). But the card everyone was waiting for — Gemini 3.5 Pro — is still missing.
Where Is Pro?
Google last updated the Pro line in February with Gemini 3.1 Pro. In the five months since, the competition has been well supplied: OpenAI shipped GPT-5.5 and began rolling out GPT-5.6; Anthropic delivered Claude Opus 4.8 and Sonnet 5; Moonshot open-sourced the 2.8-trillion-parameter Kimi K3; Zhipu released GLM 5.2.
In May, Google previewed Pro as “coming within a month.” It is now late July. Bloomberg reported that 3.5 Pro faces internal delays over missing performance targets. DeepMind’s Logan Kilpatrick responded on social media that 3.5 Pro is in testing with partners — no date attached. More intriguing still, Google formally mentioned the next flagship for the first time, saying Gemini 4 “has begun pretraining, the most ambitious pretraining effort yet.” Skipping 3.5 Pro’s formal launch while teasing generation 4’s progress is, in product-rhythm terms, an unusual signal.
The Price of Efficiency-First
Sundar Pichai said it plainly earlier this year: “Some companies burn through their whole-year token budget by May.” Aimed at enterprise customers, efficiency-first is coherent business logic. The question is whether the quality trade is worth it once models slim down this much.
Developer feedback relayed by 36Kr: Gemini 3.6 Flash repeatedly ignored instructions while building 3D scenes, outputting worse results than the 3.1 Pro released five months ago. Artificial Analysis scoring shows 3.6 Flash identical to 3.5 Flash, ranked behind Meta Spark 1.1, GLM-5.2, GPT-5.6 Luna, and Claude Sonnet 5 — no progress at all. Meanwhile Kimi K3’s explosive demand fused its compute and suspended new subscriptions within 48 hours. The market’s foot traffic says what it values today is the model’s capability ceiling, not token cost. Hence the exposed question: in a capability-is-king market, can an efficiency-first strategy offset falling behind on capability?
The Strategic Logic Underneath
Seen from another angle, Google’s choice is not unreasonable. The three models form a deliberate framework: tier one (Flash Cyber) attacks long-tail, high-value verticals with specialist models — security does not need general capability; it needs precision and trustworthiness. Tier two (Flash-Lite) covers high-throughput scenarios — search, document processing, data extraction — where latency matters and absolute reasoning depth does not. Tier three (3.6 Flash) covers most everyday tasks with efficiency gains that make enterprise adoption safe to budget. Together the tiers span 80%+ of enterprise use cases. The missing 20% — tasks demanding extreme reasoning depth — is Google’s soft spot today, and precisely the target market for 3.5 Pro and Gemini 4. In other words, Google is running “layered coverage, breakthrough by tier”: light models for user reach and data accumulation, while next-generation large-model pretraining advances quietly underneath.
What It Means for Developers
For high-throughput search or document-processing applications, Flash-Lite’s price-speed combination genuinely competes — 350 tokens per second plus adjustable “thinking levels” (low thinking for simple tasks and low latency, higher tiers for hard ones) is a more pragmatic architecture than chasing reasoning maxima. But if your application depends on deep coding reasoning or complex multi-step planning, 3.6 Flash may not be enough — at similar price points, Claude Sonnet 5 and Kimi K3 are steadier on coding benchmarks. The “local 4B does it better” jab carries emotion, and real signal.
Worth noting: 3.6 Flash’s computer-use improvement — OSWorld from 78.4% to 83% — is valuable for agent-workflow developers, and the capability is now a built-in client tool in the Gemini API and Gemini Enterprise. All three models are available from launch day in Google AI Studio, Android Studio, and the Gemini app, plus Google Antigravity and Gemini Enterprise. Figma, Harvey, Hebbia, and JetBrains offered positive early feedback on “balancing efficiency, accuracy, and speed.”
Summary
On the surface, three new models. Underneath, a clear signal: the AI industry is moving from a parameter arms race to layered efficiency competition. Not every scenario needs the best model, and not every company can afford unlimited token burn. But efficiency-first only works if you don’t fall too far behind on flagship capability — 3.5 Pro keeps slipping, Gemini 4 has just started pretraining, and the competition is already shipping substantive capability upgrades. The spectacle of this race: you bet on efficiency while the other side crushes with scale. Before the endgame, neither route is settled. One thing, though, is certain — the cover image for Google’s July story should be an unfinished draft: full of efficiency formulas, missing the final proof.
