Gemini 3.5 Flash: Google's New Default AI Model
TL;DR
Google introduced the Gemini 3.5 family at I/O 2026, starting with 3.5 Flash as the default for the Gemini app and AI Mode in Search.
May 19, 2026
Gemini 3.5 Flash GA — already the default in the Gemini app and AI Mode in Search
June 2026
Gemini 3.5 Pro target window per Google — no firmer date published
"4x faster"
Google's claim vs. other frontier models — methodology not detailed in keynote materials
$100 / $200
Gemini Spark beta scoped to U.S. Google AI Ultra subscribers on both Ultra tiers
Google introduced the Gemini 3.5 family at I/O 2026 on May 19, kicking it off with Gemini 3.5 Flash — available immediately as the default model for the Gemini app and AI Mode in Google Search. Gemini 3.5 Pro is not available yet; Google says it is coming next month. The framing across the keynote and developer day was consistent: this is a model family built for agentic workflows and coding, not just conversational chat.
What "default" actually means. Anyone using the Gemini app or AI Mode in Google Search right now is already on Gemini 3.5 Flash — no opt-in, no setting change required. That is the most concrete shift of the keynote. The same model is also available through the Gemini API, Google AI Studio, and the Managed Agents service. Gemini 3.5 Pro — which Google says outperforms Flash on the harder benchmarks — is not in any of those surfaces yet. Google says Pro arrives in June 2026; no firmer date has been published.
The agentic framing. Google describes Gemini 3.5 Flash as "the strongest agentic and coding model" in its family, and says it was co-developed using Antigravity, the company's own development platform (also updated at I/O). Google claims Flash outperforms the previous Gemini 3.1 Pro across nearly all benchmarks and runs "four times faster than other frontier models." Both claims are Google's. Independent benchmarks comparing Gemini 3.5 Flash directly against GPT-5.5 or Claude Opus 4.7 are not yet public, and the published methodology behind the "4x faster" figure was not detailed in the keynote materials. Treat both as Google's marketing positioning until third-party numbers land.
Omni and Spark. Google also previewed two related products that are not yet generally available. Gemini Omni is a multimodal generation model that produces text, image, and video output from a single prompt. Gemini Spark is a 24/7 personal agent that runs in the background and takes actions across email, calendar, and other Google services on a user's behalf. Google says Spark is in limited beta and currently only available to U.S. Google AI Ultra subscribers — on both the $100 and $200 tiers. Omni's broader availability timeline was not detailed at the keynote.
Why this matters for AI tool users. For most AI tool users and builders, the meaningful shift is not the benchmark headline — it is that Google's default tier is now an agentic model instead of a chat model. If you build agent workflows on Google AI APIs, your default model on the Gemini app or in AI Studio is now tuned for multi-step tool calls, code generation, and structured output rather than open-ended conversation. If you have been comparing Gemini against ChatGPT or Claude for serious work, the "version you are testing" has materially changed since last week. For solo builders running quick prototypes, this lands as a free upgrade. For teams deciding API providers, it warrants a fresh round of benchmarks against your actual workload — particularly if those workloads involve tool calls, multi-step reasoning, or code generation. See our roundup of best AI chatbots for where Gemini currently lands against ChatGPT and Claude, and the AI Tool Pricing Database for current Gemini API rates next to its competitors.
What to do this week. Don't switch API providers based on the keynote alone — Google's own benchmark claims still need third-party validation, and Gemini 3.5 Pro (the heavier model that may actually compete on the hardest tasks) is not yet shipping. But if you have already been testing Gemini for any workflow, run a quick agent-style task — multi-step tool calls, code generation, structured output — and see if Flash's behaviour has changed in ways that matter for what you build. If you build coding agents specifically, our Cursor review and Claude Code review compare the leading agent-first coding tools; the underlying model is a separate decision worth making explicitly. Hold any Gemini 3.5 Pro evaluation until June, when Google says it ships. If you're earlier in the stack-selection process, Find My Tool can help narrow the choices.
Why It Matters
Google's default tier is now an agentic model, not a chat model. That is the underlying shift, regardless of whether the benchmark headlines hold up to independent testing. For anyone building on the Gemini API or comparing Google's offering to ChatGPT and Claude, the baseline assumption has changed: the model behind the default surfaces is tuned for tool calls, code, and multi-step reasoning. The Gemini 3.5 Pro delay matters too — the heavier model that would compete directly with GPT-5.5 and Claude Opus 4.7 on the hardest tasks is not in anyone's hands yet. That is the gap to watch in June.
Who's Affected
- — Anyone using the Gemini app or AI Mode in Search. You are already on Gemini 3.5 Flash. No setting, no opt-in — Google rolled it out as default. If you noticed Gemini responses behave differently this week, that is why.
- — Solo builders and small teams on the Gemini API. Your default API model has materially changed. Existing prompts and agent workflows should be re-tested. The agentic positioning suggests Flash should now be stronger on tool-call reliability and structured output, but verify against your actual workload before committing.
- — Developers evaluating Gemini vs. Claude vs. GPT-5.5. The Gemini side of that comparison has shifted; benchmarks from last month are stale. Re-run them. Pro arrives in June, and that is the version that will set the real ceiling on Google's competitive position.
- — Google AI Ultra subscribers (U.S.). Gemini Spark is available in limited beta on both the $100 and $200 Ultra tiers — if you are paying for Ultra and based in the U.S., it is worth claiming a slot and seeing how a 24/7 personal agent fits into your actual workflow.
What To Do Now
- 1. Don't switch API providers on the keynote alone. Google's own benchmark claims need third-party validation, and Pro is the version that may actually move competitive positioning. Wait until June before committing to a model migration.
- 2. Re-run your agent benchmarks. If you have any agent workflow that hits the Gemini API today, run it again this week. The default model under your code changed without you doing anything — quietly behaving differently is a real failure mode.
- 3. Treat "4x faster" as marketing. Google did not detail the methodology behind that figure in keynote materials. Until the benchmarks are public, do not quote it as fact in your own decision documents.
- 4. Keep your coding-tool decision separate from your model decision. Cursor and Claude Code are tool-layer choices; the model underneath is a different lever. Don't conflate the two when evaluating what to build with this week.
More on this topic — Best AI Chatbots
Independent Review
Gemini
Pricing, pros and cons, real-world verdict — no affiliate spin.
Read the Gemini reviewMore from ToolNav News