What you need to know
- Flash has shipped, Pro has not. Gemini 3.5 Flash reached general availability on 19 May 2026. Gemini 3.5 Pro is in limited preview for select Vertex AI customers, with general availability only expected in June 2026 — Google has not committed to a date.
- Flash is the new default. It now powers the Gemini app and Google Search AI Mode, and Google markets it as roughly four times faster on output tokens per second than other frontier models.
- Flash beats last year's Pro on many benchmarks — but still trails Gemini 3.1 Pro on the hardest reasoning and the deepest long-context retrieval.
- Pro is slated to bring a 2M-token context window and stronger hard-reasoning, but those specifics are reported and expected, not confirmed shipped.
- The real signal is Omni. Google DeepMind's world-model family — Gemini Omni and Omni Flash — reframes the roadmap around any-to-any multimodality and AGI, not chat.
Gemini 3.5 Pro has not reached general availability. At Google I/O 2026 Sundar Pichai said "give us until next month" with no committed date, and the model remains a limited preview for select Vertex customers. Do not plan production migrations around unpublished Pro benchmarks or pricing — none of that is confirmed yet.
What Flash already proves
The confirmed, shipped story is Flash. Gemini 3.5 Flash went to general availability on 19 May 2026 and immediately became the default model behind the Gemini app and Google Search AI Mode — which means it is, by volume, one of the most-used frontier models on the planet from day one. Google's headline claim is speed: roughly four times the output tokens per second of competing frontier models, while still posting numbers that, on many benchmarks, beat last year's Gemini 3.1 Pro.
That is the part builders should sit with. A "Flash" tier — the cheap, fast tier — now outscores the previous generation's flagship on a broad swathe of evaluations. The frontier is not just moving up; the floor is moving up faster than the ceiling.
| Benchmark (per Google) | What it measures | Gemini 3.5 Flash |
|---|---|---|
| Terminal-Bench 2.1 | Agentic terminal / tool use | 76.2% |
| GDPval-AA | Economically valuable task quality (Elo) | 1656 Elo |
| MCP Atlas | Model Context Protocol tool orchestration | 83.6% |
| CharXiv Reasoning | Multimodal chart / figure reasoning | 84.2% |
These are Google's own published figures, not independently reproduced, so treat them the way you would any first-party benchmark — directionally useful, worth re-running on your own evaluation set before you bet a roadmap on them. The Terminal-Bench and MCP Atlas scores matter most to the agent builders among our readers: they are proxies for how well the model holds a tool-calling loop together, which is where a lot of Indian and UK product teams are spending their 2026.
Where Flash still loses to last year's Pro
The marketing line — "Flash beats 3.1 Pro" — is true on average and misleading at the edges. Flash still trails Gemini 3.1 Pro on exactly the workloads that Pro tiers exist to serve:
- Upper-end long-context retrieval. On the 128k-plus slice of MRCR (multi-round co-reference resolution), Gemini 3.1 Pro keeps roughly a 7.6-point lead over Flash. If your product genuinely reasons over very long documents, the cheap tier is not a free lunch.
- Hard reasoning. On Humanity's Last Exam and ARC-AGI-2 — the benchmarks built specifically to resist pattern-matching — 3.1 Pro remains ahead. These are the closest public proxies we have for "can it actually think through a novel problem".
For a builder, that is the whole decision in miniature. Flash is the right default for the overwhelming majority of throughput — search, summarisation, agentic tool loops, multimodal extraction. The Pro tier earns its keep only on the long tail of genuinely hard, genuinely long tasks. If you are running a careful cost-per-task analysis, our breakdown of the Flash pricing and cost economics is the companion to read alongside this one.
"We route 90% of traffic to Flash now and reserve the Pro-class tier for a single step — the long-context legal-clause cross-reference where the cheap model still drops co-references. The trap is assuming 'beats last year's Pro on average' means 'beats it on your hardest path'. It does not. Benchmark your own worst case."
— Ananya, Verified Builder · Bengaluru, INWhat Gemini 3.5 Pro is expected to add
Here is the part that demands hedged language, because almost none of it is confirmed. As of late May 2026, Gemini 3.5 Pro is a limited preview available only to select Vertex AI customers. General availability is expected in June 2026; Sundar Pichai's public framing at Google I/O was "give us until next month", which is a soft commitment at best.
The headline expected specification is context length. Google is reported to be readying Pro with a 2M-token input context window — double Flash's 1M, and, if it ships as described, the largest of any production frontier model at that time. That is the number that would matter most to long-context teams: whole-codebase reasoning, multi-document compliance review, and book-length analysis in a single pass. But until Pro is generally available with published numbers, treat the 2M figure as expected and unverified.
Do not build a product feature on the assumption of a 2M-token window before Pro is generally available. Architect your retrieval so that context length is a tunable, not a hard dependency — chunk and rank as if you had 256k, then widen the window when (and if) the larger context lands and you have verified its recall on your own data. That way a slipped general-availability date does not slip your roadmap.
The other expected gain is on hard reasoning and the upper-end long-context retrieval where Flash currently trails. That is the natural job of a Pro tier, and it is what Google has signalled. But we are not going to print benchmark numbers Google has not published. When Pro reaches general availability with first-party scores, re-run your own evaluations before trusting either the marketing or us.
| Capability | Gemini 3.1 Pro (shipped) | Gemini 3.5 Flash (shipped) | Gemini 3.5 Pro (expected) |
|---|---|---|---|
| Status | Generally available | GA — 19 May 2026 | Limited preview · GA expected Jun 2026 |
| Input context window | 1M tokens | 1M tokens | 2M tokens (reported, unconfirmed) |
| Output speed | Baseline | ~4× faster (per Google) | Not disclosed |
| Hard reasoning (HLE, ARC-AGI-2) | Leads Flash | Trails 3.1 Pro | Expected to lead (unverified) |
| Pricing | Published | Published | Unknown / expected |
Building on Gemini, Claude or GPT? Show the work.
Every article on AI Tech Connect is written by a Verified Builder. Browse profiles, shortlist who you want to hire or collaborate with — or add your own.
Become a Verified Builder →The Omni world-model bet
The chat-model horse race is the headline, but it is not the bet. At Google I/O 2026, announced on 19 May, Google DeepMind unveiled Gemini Omni and Omni Flash — a multimodal "world-model" family aimed explicitly at AGI. The pitch is any-to-any: Omni Flash can take any input modality and generate any output modality, starting with video, with image and text generation to follow.
"World model" is the load-bearing phrase. The argument inside DeepMind is that a system which has learned a predictive model of how the world behaves — physics, cause and effect, the consequences of actions across time — is a more promising path to general intelligence than a text predictor with multimodal bolt-ons. This is the same intellectual lineage as DeepMind's interactive-environment work; if you have been following the Genie 3 world-model research, Omni is the productisation of that thesis at frontier scale.
For builders, the near-term reality is more modest than the AGI framing suggests. Any-to-any generation "starting with video" means video output is the first concrete capability, and the rest is roadmap. But the strategic signal is real: Google is reorganising its frontier around multimodal world modelling, not just bigger language models. That is the lens to read the whole Gemini 3.5 generation through — Flash and Pro are the chat-shaped surface of a research programme that is aiming somewhere else entirely.
What it means for builders in India and the UK
Strip away the AGI rhetoric and three decisions land on your desk this quarter.
Cost and the new default
Flash being the default — fast, cheap, and already beating last year's flagship on most evals — changes the baseline economics for everyone routing high-volume traffic. For a Bengaluru SaaS team or a London fintech running millions of inference calls a month, the question is no longer "Flash or Pro" but "what is the smallest slice of traffic that genuinely needs a Pro-class tier". For most products that slice is small, and it shrinks every generation.
Context length as a moat — maybe
If Pro ships with a verified 2M-token window, it becomes the obvious choice for whole-repository reasoning, multi-document compliance review, and long-horizon agent memory — the workloads where Claude and GPT currently compete hard. But "if" is doing real work in that sentence. Until general availability and independent recall testing, treat the 2M window as a reason to keep Gemini on your evaluation shortlist, not a reason to migrate.
Multimodal and agents
Flash's strong showing on Terminal-Bench and MCP Atlas, plus the Omni any-to-any direction, makes Gemini a serious contender for agentic and multimodal pipelines specifically. If your product is computer-use, tool-orchestration, or multimodal extraction, Gemini deserves a head-to-head against the field — our running comparison of Claude, OpenAI Codex and Gemini for computer-use agents tracks exactly that race. And for the research-minded, DeepMind's track record of turning frontier models into genuine discovery tools — see AlphaEvolve's algorithm-discovery work — is part of why the Omni bet is worth taking seriously rather than dismissing as marketing.
The honest summary
Google's frontier in June 2026 is a split picture. Flash is shipped, fast, cheap, and genuinely strong — the part you can act on today. Pro is the headliner that has not arrived, carrying expected gains in context length and hard reasoning that nobody outside select Vertex customers has verified. And Omni is the long bet — a world-model family that reframes the whole programme around AGI rather than chat.
The builder's move is unglamorous: default to Flash, keep Pro on the shortlist for the hardest long-context work, and re-run your own evaluations the moment Pro reaches general availability. Believe the shipped benchmarks cautiously, the unshipped ones not at all, and the AGI framing as a statement of direction rather than a delivered capability.