Gemini 3.5 Pro Launches: Specs, Deep Think, Pricing

After two slipped targets and a base-model rebuild, the wait is over. On July 17, 2026, Google DeepMind released Gemini 3.5 Pro, its most capable model to date and the new flagship of the Gemini 3.5 family. The release confirms the specifications that had circulated for weeks as leaks — a 2 million-token context window, a Deep Think reasoning mode reserved for the top subscription tier, and pricing that plants the model firmly at frontier rates. It arrives on the same day the Shanghai World AI Conference opens, a timing that puts Google’s answer to a fast-moving frontier squarely in front of a global audience.

We previewed this launch last week when the July 17 date and the headline specs were still unconfirmed rumor. With the model now shipping, here is what Google has actually documented.

What shipped

Gemini 3.5 Pro replaces the Ultra tier as the capability ceiling of Google’s lineup, sitting above the already-shipped Gemini 3 family. The confirmed headline features:

• A 2 million-token context window. This is the specification that matters most in practice — double the 1 million tokens that had been standard on Google’s Pro tier and on most rivals, including Anthropic’s Claude line. It lets the model hold entire books, large codebases, or long agent histories in a single request. Our primer on what a context window is explains why the number translates directly into what developers can build.
• A Deep Think reasoning mode. Gemini 3.5 Pro ships a dedicated Deep Think layer aimed at logic, mathematics, and multi-step problems — the “think before answering” approach behind the broader class of reasoning models. Crucially, Deep Think is gated to the $250-per-month Ultra subscription, not the $20 Pro consumer tier, a segmentation that reserves the model’s hardest reasoning for its highest-paying users.
• Multimodal text and image support, consistent with the rest of the Gemini 3.5 line.
• Stronger coding and long-horizon reasoning, the two areas Google has said the rebuilt base model was specifically retrained to fix.

Prior to today, Gemini 3.5 Pro had been available only in a limited enterprise preview through Vertex AI for select partners. General availability opens it to the broader developer base through the Gemini API and Vertex AI.

The pricing

Google set API pricing at roughly $15 per million input tokens and $60 per million output tokens — about ten times the cost of Gemini 3.5 Flash, the speed-tier model that now powers Google’s AI Mode in Search. Those rates sit well above the budget tiers and at the upper end of the frontier band, confirming that Google intends to charge premium prices for premium capability rather than compete on cost at the top of its lineup.

The structure mirrors a barbell strategy that now defines the whole field: keep a fast, cheap model (Flash) for scale and everyday consumer surfaces, and charge frontier rates for the Pro flagship that enterprises license and build products on. Deep Think’s confinement to the $250 Ultra tier is the sharpest expression of that segmentation — the most expensive reasoning is available only to the users willing to pay the most for it.

Why the delay mattered

Gemini 3.5 Pro was announced at Google I/O on May 19, 2026 with a June general-availability target. That date slipped, then slipped again. Reporting through the delay described early enterprise testers flagging gaps in reasoning and coding during the limited preview, and ultimately a decision to scrap the original base architecture in favor of a fresh pretraining run — a change substantial enough to explain a multi-week slide and the “rebuild” framing that accompanied it.

Two specific failure modes were reported in the scrapped build: it could not maintain structural consistency when generating complex, multi-layered SVG scene layouts, and it broke down under recursive tool-calling environments — precisely the agentic workloads that matter most to the developers Google is courting. Whether the rebuilt model closes those gaps is now a question the market can test directly, rather than infer from leaks.

The strategic reading is straightforward. Google is launching into a frontier that reset sharply in early July. OpenAI shipped its GPT-5.6 family — Sol, Terra, and Luna, and xAI released Grok 4.5, pitched as an “Opus-class” coding model at aggressive prices. Shipping a Pro flagship that underperformed those on coding and reasoning would have been a costly unforced error. Google’s willingness to be late twice signals it would rather ship a competitive model than an on-time one.

The benchmark question

Here the picture is still forming. During the preview period, early-tester reporting suggested the pre-rebuild Pro model trailed Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 on advanced reasoning, coding, and long-horizon task execution — the very weaknesses the rebuild was meant to address. Whether the shipped model closes that gap is the first thing enterprise developers will measure against their own workloads.

The sensible posture is to treat the launch benchmarks as a starting point, not a verdict. Frontier-model leaderboards move quickly, and the meaningful test for a Pro-tier model is not a single score but reproducible performance on the coding and agentic tasks that paying developers actually run. The 2M-token window is the one differentiator that does not require a benchmark to evaluate: for whole-repository analysis, large document sets, and durable agent memory, a doubled context window is a capability developers can feel immediately.

Where it fits in the race

Gemini’s distinctive advantage has never been a single benchmark; it is distribution. Google can drop a model straight into Search and a billion-plus everyday users — exactly what Gemini 3.5 Flash already does in AI Mode. A stronger Pro tier feeds the top of that funnel: the model enterprises license through Vertex AI, the one whose capabilities eventually cascade down into the Flash models serving consumers.

On the capability axis, 3.5 Pro will be measured head-to-head against OpenAI’s GPT-5.6 Sol and Anthropic’s most capable Claude models, with coding and agentic performance the battlegrounds that matter most to paying developers. On economics, the confirmed pricing shows Google is comfortable charging frontier rates for Pro while keeping Flash cheap for scale — the same barbell every major lab now runs.

What it means

The launch matters less as a single release than as Google’s confirmation that it will not cede the frontier. With GPT-5.6 and Grok 4.5 already shipping and Chinese labs closing ground, a competitive Pro flagship was table stakes. Google chose to slip the date twice rather than ship something that underwhelmed — a defensible trade when the competitive set had just reset expectations.

Who wins. Enterprise developers on Vertex AI gain the most: a 2M-token window and stronger agentic tool use directly expand what they can build and ship. Google’s Search and consumer surfaces are the quieter beneficiary, since a better Pro tier eventually improves the Flash models everyday users touch. Anthropic and OpenAI now face a credible third frontier competitor with distribution neither can match.

Who’s exposed. The $250 Ultra gate on Deep Think is a bet that the hardest reasoning is worth a premium subscription — a bet that only pays off if the model’s reasoning is demonstrably ahead of cheaper rivals. If independent benchmarks show 3.5 Pro merely matching GPT-5.6 and Claude rather than beating them, the pricing looks aggressive rather than justified.

What to watch next. Three things separate a strong launch from a loud one: independent benchmark reproductions on real coding and agentic workloads, developer migration from rival APIs now that the 2M window is generally available, and whether the promised improvements in recursive tool use hold up in production agent stacks. The specs are no longer leaks. Now they have to perform.