Google DeepMind’s Gemini 3.5 Pro is targeting a July 17, 2026 launch, according to leaked details that surfaced in early July. The model reportedly ships with a 2 million token context window — double anything currently in the frontier field — and a new “Deep Think” reasoning layer for multi-step problem-solving. But the leaked plans also suggest Deep Think will be gated behind Google’s $250/month Ultra subscription, raising the question of whether frontier reasoning capability is becoming a luxury good.
🔍 THE BOTTOM LINE
If the leaks hold, Gemini 3.5 Pro lands five days after GPT-5.6 and nine days after Grok 4.5 — the most competitive single week in AI history. A 2 million token context window is genuinely new. But locking the reasoning mode behind a $250/month paywall means the model’s most differentiated feature is unavailable to most developers. Google is betting that capability tiering, not capability itself, is the path to revenue.
What the Leaks Say
The details trace back to a YouTube breakdown by the channel World of AI, later summarised by Geeky Gadgets on July 6. Neither Google nor DeepMind has officially confirmed the feature set, naming, or launch date. With that caveat, here is what is circulating:
- 2 million token context window — roughly enough for a lengthy technical manual, a full book, or thousands of lines of code in a single pass. This doubles the current frontier, though advertised window size does not always translate to equally strong recall across the full span.
- Deep Think Reasoning Layer — a dedicated reasoning stage for multi-step logic, math, and decision-making. Similar in concept to OpenAI’s o1-style extended reasoning, but reportedly built directly into the model rather than layered on top.
- Autonomous workflows for coding, tool use, and task execution.
- SVG generation and front-end design output positioned as a differentiator.
- 3D modelling and visual coding capability.
- A phased rollout tying Gemini 3.5 Pro to future releases like Gemini 3.6.
TechTimes reported on July 8 that the July 17 target is firm, with DeepSeek’s own July 24 deadline for developers creating additional competitive pressure.
The Pricing Question
The most strategically significant detail is not the context window — it is the paywall. Build Fast With AI’s July 12 roundup reports that Deep Think will be gated behind the $250/month Google AI Ultra subscription tier, with expected API pricing around $1.25 input and $10 output per million tokens for the base model.
That API pricing would undercut GPT-5.6 Sol ($5/$30) by 4x on input cost while doubling the context window. But the reasoning mode — the feature most likely to matter for complex coding, research, and analysis — sits behind a consumer subscription, not an API tier. This creates a two-tier model: developers get the context window via API, but the reasoning layer is reserved for subscribers.
Google’s own subscription page lists Google AI Ultra at $99.99/month for 5x higher usage limits and $199.99/month for 20x higher limits. The $250 tier reported in the leak would be a new premium above both existing plans. If accurate, Google is creating a three-tier access structure: free (limited), Pro ($20), Ultra ($100-$200), and the new top tier ($250) — each buying progressively more reasoning capability.
The Competitive Position
Gemini 3.5 Pro arrives six weeks behind its originally expected window. As we reported in June, the delay raised questions about Google’s competitive edge in coding benchmarks. The intervening weeks were not kind: Google lost Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic, while GPT-5.6 launched publicly and Grok 4.5 landed as the fastest coding model at half the price.
The 2 million token context window is the one spec that no competitor matches. GPT-5.6 Sol offers 400K tokens. Claude Fable 5 offers 500K. GLM-5.2 offers 1 million. If Gemini 3.5 Pro’s 2 million tokens perform well across the full span — not just in the first 128K — it becomes the default model for large-document tasks, codebase analysis, and multi-hour transcripts. Kalinga AI’s leak analysis notes that “a large advertised window doesn’t always translate into equally strong recall across the entire span.”
The real test will be Terminal-Bench 2.1. GPT-5.6 Sol posts 88.8% standard and 91.9% in its Ultra configuration. If Gemini 3.5 Pro with Deep Think enabled can match or exceed that, the paywall argument gets harder to criticise. If it cannot, Google is selling a large context window with a reasoning mode that underperforms — at a premium.
Why the Paywall Matters
The trend across frontier labs is toward tiering capability by price. OpenAI gates its best reasoning behind higher API costs. Anthropic gates its best model behind higher API costs and limited availability. Google is now gating reasoning behind a subscription. The pattern is clear: the models that can think hardest cost the most, and the gap between “good enough” and “best available” is widening in dollar terms.
This is the context in which Zhipu’s open-weights GLM-5.2 and Cognition’s SWE-1.7 matter. Both offer near-frontier capability at a fraction of the cost — or, in GLM-5.2’s case, free. If Google puts Deep Think behind a $250/month paywall, the open-weights argument gets louder: why pay $250/month for reasoning when a MIT-licensed model offers comparable agentic performance at no cost?
NZ Angle
New Zealand developers building AI-powered applications face the same tiering decision as everyone else, but with an added exchange-rate penalty. A $250/month subscription is roughly NZD $410 at current rates. For a solo developer or small startup, that is a meaningful line item — especially when the alternative is an open-weight model that runs locally on consumer hardware. The Gemini 3.5 Pro leak, if accurate, means the cost of frontier reasoning capability in New Zealand just went up.
❓ FAQ
Is the July 17 launch date confirmed? No. The date comes from leaked details first reported by the YouTube channel World of AI and summarised by Geeky Gadgets on July 6. Google has not officially confirmed the date, the feature set, or the pricing. Treat all specifics as rumour until Google’s announcement.
What is Deep Think? A dedicated reasoning layer built into Gemini 3.5 Pro for multi-step logic, math, and decision-making. It is similar in concept to OpenAI’s o1-style extended reasoning but reportedly integrated into the model architecture rather than layered on top. The leak suggests it will be gated to the highest subscription tier.
How does 2 million tokens compare to competitors? GPT-5.6 Sol: 400K tokens. Claude Fable 5: 500K tokens. GLM-5.2: 1 million tokens. Gemini 3.5 Pro’s reported 2 million would be the largest available, though real-world recall across the full span remains unverified until independent benchmarks appear.
Should developers wait for Gemini 3.5 Pro or use GPT-5.6 now? The answer depends on use case. If you need large-context document analysis, waiting for 2 million tokens may be worth it. If you need coding performance now, GPT-5.6 Sol is already available and benchmarked. The Deep Think paywall means the full reasoning capability may not be accessible via API at any price.
Why is Google releasing this six weeks late? The model was originally expected in late May or early June. Google has not explained the delay. The intervening period saw key talent departures (Shazeer to OpenAI, Jumper to Anthropic) and competitor launches (GPT-5.6, Grok 4.5). The delay may reflect a rebuild rather than a polish — the leak suggests an entirely new pretraining run, not an adaptation of 2.5 Pro.
🔍 THE BOTTOM LINE
Gemini 3.5 Pro’s leaked specs are aggressive on paper: 2 million tokens, a reasoning layer, and API pricing that undercuts the field. But the paywall on Deep Think signals where frontier AI is heading — not toward democratisation, but toward a pricing structure where the ability to think deeply is a premium feature. The open-weights alternative just became more strategically important.
