Google's Gemini 3.6 Flash is cheaper and shorter - and Gemini 4 gets a tease while 3.5 Pro stays late

Google publie trois nouveaux modèles Gemini avec un focus « moins de tokens en sortie, moins cher », teste Gemini 4 dans la salle d'attente - et laisse Gemini 3.5 Pro sur la ligne de départ.

In plain terms - Google shipped three new Gemini variants on 21 July 2026, headlined by Gemini 3.6 Flash. The pitch is straightforward: fewer output tokens, lower prices, and a cybersecurity-tuned model in the mix. But Gemini 3.5 Pro - the flagship reasoning model teased for weeks - is still in testing. Google is instead pre-announcing Gemini 4.

Contexte

Google's Gemini cadence in 2026 has been rough: 3.5 Pro slipped multiple times, with Google privately citing coding-benchmark regressions. The launch-what-is-ready-now approach (Flash, plus adjacent specialised variants) is a way to keep the release drumbeat going while the reasoning flagship is retooled.

Les données rapportées

• Three models announced: Gemini 3.6 Flash (general purpose, faster), a cybersecurity-oriented variant, and a third undisclosed variant on the way to public preview.
• Google claims reduced output tokens for equivalent tasks vs 3.5 Flash, translating to lower per-request cost. Exact benchmark deltas not published in the launch post.
• Gemini 3.5 Pro is confirmed still in testing.
• Gemini 4 is teased - no date, no capability spec, no pricing.

Analyse

Two things are happening at once. First, Google is committing to a two-tier model reality: Flash-class models optimised for cost and latency, Pro/Ultra-class for hardest reasoning. That mirrors OpenAI's GPT-5.6 vs Codex/Work split and Anthropic's Fable vs Haiku split. The three frontier labs are converging on the same product topology.

Second, the Gemini 4 tease while 3.5 Pro isn't out is a signal, not a leak. It is Google telling markets and enterprise buyers: the reasoning ceiling is still moving, don't lock in on competitors. The bet is that "there's more coming" is enough to keep the pipeline warm even when the flagship is late.

For engineers on the buy-side: 3.6 Flash's real value is measurable - output-token reductions compound in production. Test it against your worst latency tasks. For the roadmap conversation with a CFO: don't sign multi-year Gemini contracts until 3.5 Pro is actually GA.

Scénarios

• Base case (55%) : Gemini 3.5 Pro ships within 60 days at parity or slight lead vs Fable 5 on reasoning; Gemini 4 lands Q1 2027.
• Slip (30%) : 3.5 Pro slips again; Google leans harder on Flash tier while Anthropic and OpenAI extend their reasoning lead.
• Leapfrog (15%) : Google skips 3.5 Pro entirely and jumps to 4 - the "Windows 9" playbook.

Implications

For a builder: don't wait - 3.6 Flash is worth benchmarking now. For a decision-maker: the two-tier reality is here; buy the tier you actually need, not the flagship you'd like to name-drop.