Google Delays Gemini 3.5 Pro Past Its I/O Promise

"Give us until next month," Sundar Pichai told developers at Google I/O on May 19, 2026, when pressed about when Gemini 3.5 Pro would reach general availability. The audience audibly groaned.

One month passed. Then another week. As of June 27, Gemini 3.5 Pro still isn't publicly available.

TL;DR

Pichai promised Gemini 3.5 Pro general availability by the end of June at Google I/O 2026
Six weeks later: limited Vertex AI enterprise preview only, no public API, no benchmark release
Google has "declined to comment" on the revised timeline; July is the current expectation, not a commitment
Four senior Gemini researchers left for Anthropic and OpenAI the same week the deadline passed

The Claim

Sundar Pichai introduced Gemini 3.5 Pro to a packed auditorium at the Shoreline Amphitheatre on May 19. He described the model as already being used productively inside Google's own teams, and when asked about availability he made his ask simple: wait one more month. That same day, Google shipped Gemini 3.5 Flash to general availability - a model that already beats the prior-generation Pro on several agentic benchmarks.

The positioning was deliberate. Flash handles volume workloads. Pro handles the hardest problems. The two would form a complete lineup.

What Google specifically promised for Pro was concrete: a 2 million token context window, a reasoning mode called "Deep Think" designed for multi-step planning, and multimodal inputs. The model would also absorb use cases previously routed to Google's Ultra tier, consolidating the product line around two models instead of three.

June was the month. Then June came and went.

The Evidence

What's Actually Available

As of June 27, Gemini 3.5 Pro remains in limited enterprise preview on Vertex AI. There's no public model ID, no API access for independent developers, no published benchmarks, and no pricing sheet. Google has given a narrow set of enterprise customers access for testing, but the full scope of that program hasn't been disclosed.

A Google spokesperson "declined to comment" when reporters asked about the revised timeline, according to multiple independent accounts. No official announcement about the delay has been published.

Prediction markets assigned a 50-55% probability to a June launch at the start of the month. By June 27, those odds had collapsed.

What This Costs Developers

Several enterprise teams planned their Q3 architecture decisions around a June availability date. The 2 million token context window was a specific reason - it changes how teams handle long-document retrieval, extended agent sessions, and large-codebase analysis. Without a firm date, those decisions sit in limbo.

The interim option most evaluators recommend is Gemini 3.5 Flash for agentic and tool-use workloads, where it's already in GA with strong performance. For long-context retrieval specifically, GPT-5.5 and Claude Opus 4.7 are the current alternatives. Neither offers a 2 million token context, but both have published benchmarks that make planning possible.

Claim vs. Reality

Claim	Reality
General availability in June 2026	Limited enterprise preview as of June 27
"Give us until next month"	Google has declined to announce a new date
2M token context window	Target, not confirmed in any released spec
Replaces Ultra tier	No GA model to confirm or test
Already in use internally at Google	No independent verification possible

A speaker addresses an audience at a large tech conference, viewed from behind the crowd Google's I/O keynotes have become the primary venue for model launch promises - and the primary venue for tracking whether those promises hold. Source: pexels.com

What They Left Out

The Researchers Who Built It

Between June 21 and June 27, four senior Gemini researchers announced they were leaving Google for rivals. We covered the departures last week: Noam Shazeer, co-author of the original transformer paper, left for OpenAI on June 18. John Jumper, the 2024 Nobel laureate who led AlphaFold, announced his move to Anthropic on June 19. Two more contributors - Jonas Adler and Alexander Pritzel - followed, also to Anthropic, on June 24. These weren't peripheral figures. They helped design the architecture Gemini 3.5 runs on.

Google DeepMind CEO Demis Hassabis publicly called the departures "expected." Alphabet's stock closed at $343.71 on June 24, down from its all-time high of $402.38 set on May 13 - a drop of roughly 15% in six weeks.

This Isn't the First Delay of 2026

Reporting from Bind AI notes this is Google's second major AI delivery miss of 2026, following a three-month slip with Gemini Ultra 1.5 earlier in the year. The Ultra 1.5 delay attracted less scrutiny because the competitive environment was quieter. This one lands at a worse moment: Claude Opus 4.7 is already shipping, and GPT-5.5 is in active API use. Every week without a Pro GA date is a week competitors hold the field.

How 3.5 Flash Sets the Bar

Since Pro's performance remains unverifiable without GA access, the most useful baseline is Gemini 3.5 Flash, which shipped May 19 with a full model card. If Pro delivers beyond Flash on the tasks where Flash already leads, the table below shows what it needs to clear.

Benchmark	3.5 Flash	3.1 Pro	Claude Opus 4.7	GPT-5.5
Terminal-Bench 2.1	76.2%	70.3%	66.1%	78.2%
SWE-Bench Pro	55.1%	54.2%	64.3%	58.6%
MCP Atlas (tool-use)	83.6%	78.2%	79.1%	75.3%
Finance Agent v2	57.9%	43.0%	51.5%	51.8%
ARC-AGI-2	72.1%	77.1%	75.8%	84.6%
MRCR v2 at 1M tokens	26.6%	26.3%	-	-

Source: Google DeepMind Gemini 3.5 Flash model card, May 2026

On tool use and short-context tasks, Flash is already competitive with Claude Opus 4.7 and ahead of GPT-5.5 on MCP Atlas and Finance Agent workloads. But the long-context numbers are a different story. At 1 million tokens on MRCR v2, Flash scores 26.6%. GPT-5.5 sits at 94.8%. That gap is exactly where 3.5 Pro's 2 million token context window is meant to matter. Without a GA release and independent benchmarks, there's no way to know if Pro actually closes it.

The Google I/O 2026 collection of announcements, including the Gemini 3.5 generation, shown as a Google promotional banner Google I/O 2026 brought a full slate of product announcements, but the flagship model of the Gemini 3.5 line didn't ship the same day. Source: blog.google

Google's position on Gemini 3.5 Pro as of June 28 is silence. No revised date, no benchmark preview, no confirmed pricing. Developers who planned around a June launch have three realistic options: wait for a July window that Google hasn't confirmed, route their longest-context workloads to GPT-5.5 in the meantime, or continue with Gemini 3.5 Flash and accept its current limitations. The 2 million token context window is worth waiting for, if it performs as described. The Gemini 3.5 Pro model card will be updated once GA access is available and benchmarks can be independently verified.

Sources: