Gemini 3.5 Pro

Overview

Google DeepMind announced Gemini 3.5 Pro at Google I/O on May 19, 2026, positioning it as the new flagship of the Gemini 3.5 family and the replacement for the Ultra tier in Google's model lineup. As of late June 2026, the model hasn't publicly launched - Google confirmed a delay pushing general availability to July, with current access limited to a Vertex AI enterprise preview for select partners.

What's known comes mostly from Google I/O announcements and Vertex AI preview documentation. Sundar Pichai called it "a game-changer for us internally at Google" during the keynote. The model is multimodal, supports text and image inputs, includes a Deep Think reasoning mode for multi-step problem solving, and ships with the largest context window in the Gemini family at 2 million tokens - double the 1M window on Gemini 3.1 Pro and Gemini 3.5 Flash.

In Google's lineup, Gemini 3.5 Pro replaces the Ultra tier, signaling that Google is consolidating its naming around the Pro/Flash/Flash Lite structure used in the 3.1 generation rather than maintaining a separate Ultra tier. The smaller sibling, Gemini 3.5 Flash, launched on May 19 and is already in general availability. Pro is the delayed half of the pair.

Key Specifications

Important caveats: Pricing figures are community estimates based on I/O preview materials and Vertex AI documentation leaks. Google hasn't published official API pricing for Gemini 3.5 Pro. Treat any pricing figure as speculative until the public launch.

Benchmark Performance

Google hasn't published a full benchmark suite for Gemini 3.5 Pro. The figures below come from the I/O 2026 announcement for the Gemini 3.5 family and from Vertex AI enterprise preview documentation. Competitor figures are from the same evaluation periods where available.

The benchmark column for Gemini 3.5 Pro is intentionally empty - no verified scores have been published for the Pro model as of June 28, 2026. The sibling Flash model numbers are included to give a floor: if Flash already scores 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas, the Pro tier - which Google describes as substantially more capable on complex reasoning - should post materially higher numbers on those tasks. But "should" isn't a benchmark. This table will be updated on public launch.

For current frontier model rankings, see the agentic AI benchmarks leaderboard and the coding benchmarks leaderboard, where verified scores are tracked as evaluations come in.

Key Capabilities

Deep Think Reasoning Mode

The headline differentiator for Gemini 3.5 Pro is Deep Think - a reasoning mode that allocates extended internal compute to multi-step problems before producing an answer. Google described it at I/O as a step beyond the dynamic thinking available in Flash, designed for tasks where accuracy matters more than latency: research synthesis, complex code architecture decisions, mathematical proofs, and long-horizon planning.

Deep Think isn't always-on by default. Based on Vertex AI preview documentation, it's invocable per request, similar to how extended thinking works in Claude Opus 4.7 or the "thinking budget" parameter in GPT-5.5. The distinction from Flash's configurable thinking levels is one of degree - Flash supports minimal/low/medium/high; Pro's Deep Think is positioned as a distinct, heavier reasoning mode for the hardest problems.

2 Million Token Context Window

The 2M-token window is the biggest context in the Gemini 3.5 lineup and, as of mid-2026, one of the largest available in any frontier model. To put that in scale: 2 million tokens holds roughly 1,500 pages of code, several full novels, or an entire enterprise codebase with documentation alongside it.

The practical question - how well does it actually use that context at the far end - remains open. Gemini 3.1 Pro showed long-context degradation in independent testing beyond 120-150K tokens despite its 1M window. Whether Gemini 3.5 Pro improves on that pattern is something that can only be answered after the public launch and third-party evaluation. Context window specifications and real-world usable context are different numbers; treat the 2M figure as a ceiling, not a guarantee.

Multimodal Processing

Google confirmed text and image input support at I/O. Additional modalities - audio, video, PDF - are likely given the pattern established by Gemini 3.1 Pro and 3.5 Flash, but have not been explicitly confirmed for the Pro tier. The model card will be updated with verified modality support on launch.

Pricing and Availability

Gemini 3.5 Pro is currently available only through a limited enterprise preview on Vertex AI. General public API access is not yet available as of June 28, 2026.

Google confirmed at I/O that the model was delayed from the original target of late June to sometime in July 2026. No specific launch date has been announced.

Estimated pricing (community estimates, not confirmed by Google):

For comparison, Gemini 3.1 Pro's current confirmed pricing sits at $2.00/M input and $12.00/M output at standard tier. The estimated Pro premium of 50% on input and similar on output aligns with Google's historical pro-to-standard pricing gap, but this is extrapolation, not documentation.

Competitors at similar capability tier: Claude Opus 4.7 at $5.00/$25.00 per M tokens (input/output), GPT-5.5 at $5.00/$30.00 per M tokens. If the community estimates hold, Gemini 3.5 Pro would undercut both on price while matching or exceeding them on context window.

Enterprise buyers in the Vertex AI preview have access now; everyone else waits. The Gemini app is expected to feature the model on AI Ultra and Pro plans after general release.

Strengths

• Largest context window confirmed for the family. 2M tokens is double what Gemini 3.1 Pro offered, and meaningful for enterprise document processing and codebase analysis
• Deep Think reasoning mode. Targeted at multi-step complex reasoning where Flash's configurable thinking doesn't go far enough
• Likely strong benchmark performance. If the sibling Flash model already tops some agentic benchmarks, a full Pro tier above it should post competitive frontier numbers
• Google ecosystem integration. Vertex AI, Gemini Enterprise Agent Platform, NotebookLM, and Gemini CLI are expected to adopt it after GA

Weaknesses

• Not released yet. As of June 28, 2026, the model is delayed and unavailable for public API use. No production workloads can rely on it
• No verified benchmarks. Every performance claim is from Google's own keynote previews or pre-release briefings - no independent evaluation has been published
• Pricing unconfirmed. Community estimates exist but the actual API pricing is unknown until launch
• History of long-context degradation. Gemini 3.1 Pro showed real accuracy drops beyond 120-150K tokens; whether 3.5 Pro improves this is unverified
• Preview stability unknowns. Gemini 3.1 Pro had capacity and quota issues during its own preview rollout. Enterprise preview users for 3.5 Pro face similar unknowns

Related Coverage

• Gemini 3.5 Flash - the released sibling model, available now
• Gemini 3.1 Pro - the current flagship, being replaced by this model
• Gemini 3.5 Flash Review: When Flash Surpasses Pro - our hands-on evaluation of the Flash tier
• Gemini 3.5 Flash: Real Speed, Selective Benchmarks - benchmark analysis of the Flash sibling
• Agentic AI Benchmarks Leaderboard - live frontier model rankings on agentic tasks
• Cost Efficiency Leaderboard - pricing comparisons when confirmed data is available

FAQ

When will Gemini 3.5 Pro be publicly available?

Google delayed the launch to July 2026. No specific date has been announced as of June 28, 2026. Enterprise preview access is available through Vertex AI for select partners.

What is the context window for Gemini 3.5 Pro?

2 million input tokens, confirmed at Google I/O 2026. Maximum output tokens haven't been officially specified before launch.

How does Gemini 3.5 Pro differ from Gemini 3.5 Flash?

Gemini 3.5 Pro targets complex reasoning tasks with its Deep Think mode and a 2M-token context window. Flash is optimized for speed (289 tok/s) and agentic throughput at a lower cost. They're complementary tiers, not direct substitutes.

What is the pricing for Gemini 3.5 Pro?

No official pricing has been published. Community estimates suggest around $3.00/M input and $18.00/M output tokens, but these are unconfirmed.

Does Gemini 3.5 Pro support image input?

Yes, image input is confirmed. Other modalities (audio, video, PDF) weren't explicitly confirmed in I/O announcements.

Sources:

• Gemini 3.5: frontier intelligence with action - Google Blog
• Google I/O 2026 Keynote - Google Events
• Gemini 3.5 Flash Model Card - Google DeepMind
• Gemini 3.5 Flash - Google DeepMind
• Gemini 3.5 Flash: Benchmarks, Thinking and API Guide 2026 - DigitalApplied
• Google Introduces Gemini 3.5 Flash at I/O 2026 - MarkTechPost