Gemini shuffle: Flash becomes the default, 3.5 Pro announced for June

Google is reshuffling its Gemini model lineup — and doing it at a pace that is hard to track for teams running a model router. In mid-June 2026, several sources reported that Google had made the Flash line the default model across its Gemini products: fast, cheap, built for the bulk of requests. In parallel, Gemini 3.5 Pro is announced as the next flagship — for June 2026, but without a fixed date. At Google I/O in late May, CEO Sundar Pichai only confirmed „next month”, which reportedly drew audible groans from the crowd. For the boostN audience, the relevant signal is less the individual model than the cadence: anyone running a multi-provider strategy has to revisit their default selection again.

What applied before

Google’s Gemini lineup had been in flux but with clear role assignments: a high-capability Pro model for complex reasoning and coding, a Flash model for fast, cheap bulk requests, plus Lite variants for high throughput. Anyone wiring Gemini into a router picked the model per task explicitly — and the default behavior of the end products (the Gemini app, AI Studio) was secondary to the API integration.

With the release of Gemini 3.5 Flash on May 19, 2026, that balance shifted. Per Artificial Analysis, the model reaches an Intelligence Index of 55 — nine points above its predecessor Gemini 3 Flash — while delivering over 280 output tokens per second, roughly 70 percent faster than Gemini 3 Flash and comparable models such as GPT-5.4 mini. On agentic benchmarks (Terminal-Bench, Finance Agent), Flash reportedly pulled ahead of Gemini 3.1 Pro, but lost ground on pure reasoning and long-context tasks. That made a cheap Flash model suddenly good enough for many production workloads — the precondition for promoting it to default.

What applies now

1. Flash becomes the standard choice. By making the Flash line the default across the Gemini products, Google signals that the fast, cheap model is enough for the majority of requests. Flash-tier pricing reportedly sits at around $1.50 input / $9.00 output per million tokens — an order of magnitude that makes bulk workloads predictable. Anyone who had been routing everything to a Pro model by default is potentially paying for capability they don’t need.

2. Gemini 3.5 Pro is announced, but not here. At I/O in late May, Google confirmed the Pro variant for „next month”, that is June 2026, without a model card, benchmarks, or price list. Prediction markets such as Polymarket price a release by June 30 at around 60 percent — a betting probability, not an official commitment. Until the model actually ships, every Pro specification remains an expectation.

3. The Pro figures are ambitious — and expensive. Per reporting, Gemini 3.5 Pro targets a 2M-token context window, a „Deep Think” reasoning mode, and frontier multimodality. The expected pricing of around $15 input / $60 output per million tokens is about ten times that of Flash. Pro thereby takes over the role Google previously assigned to its Ultra tier: the hardest reasoning, multimodal, and long-context tasks. All of these numbers are announced or reported, not officially confirmed.

The bigger picture

The real story is not a single model but the cadence. Within a few weeks, Google shipped Gemini 3.5 Flash, promoted the Flash line to default, and announced the next Pro model — while the previous generations are still embedded in the products. For teams that address Gemini through a model router, that means recurring maintenance: check the default model, update price assumptions, re-run eval suites. A default switch on the provider’s side can move your own cost and quality curve without you changing a single line of code.

Caution is warranted on the Pro figures. „2M tokens”, „Deep Think”, and the $15/$60 pricing come from reporting and vendor announcements, not from a published model card. As long as Gemini 3.5 Pro has not shipped, these are planning figures, not solid facts — the history of past frontier launches shows that context windows, prices, and availability dates can shift right up to the GA day. Anyone designing architecture around the 2M window now is building on an announcement.

The underlying logic is also notable: a cheap, fast Flash as the default plus an expensive, context-heavy Pro at the top is exactly the two-tier structure that OpenAI and Anthropic also run. The competition is converging on the same pattern — which makes multi-provider strategies easier to compare, but the per-provider default choice all the more important.

What you can do now

Check your Gemini default in the router. If your integration relies on an implicit default model, verify which model Google now serves — and whether it matches your latency and quality assumptions. An unnoticed default switch can change cost and response behavior.

Route deliberately by task, not blanket-to-Pro. For bulk requests (classification, extraction, short answers), Flash is enough at a fraction of the expected Pro price. Reserve expensive Pro calls for the tasks that genuinely need long context or deep reasoning.

Treat Gemini 3.5 Pro as a planning assumption, not a release date. Don’t commit production architecture to the 2M-token window or „Deep Think” before the model is available and an official model card exists. Prepare an eval run you can fire against your real tasks on GA day.

Entdecke mehr

Blog

Three Days with Claude Fable 5 — and Why Opus Stays My Daily Driver

Fable 5 fixed a bug in one shot that Opus failed at twice. I'm still not switching. My honest impressions from the trenches after three days.

Glossar

Gemini 3 Pro

Gemini 3 Pro (November 2025) was Google's third flagship model with a 1M-token context window and native multimodality — now superseded by Gemini 3.1 Pro.

Lexikon

Comparing AI models — who builds what and how to choose

The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.

Themenuebersicht

KI-ModelleGemini