Google's most anticipated model of the summer is reportedly three days out, and the calendar could hardly be more loaded. Gemini 3.5 Pro is expected to launch Thursday, July 17, the very morning Shanghai's 2026 World Artificial Intelligence Conference opens with President Xi Jinping attending in person for the first time since the event began in 2018. On a single date, the West's marquee frontier release is expected to go live while the East's most powerful leader steps onto the world's biggest AI-governance stage.
Everything specific about the launch remains, for now, reported expectation rather than confirmed fact. Google has not officially announced a date, a price, or a spec sheet. What is circulating comes from leaked plans and third-party reporting: a 2-million-token context window, a Deep Think reasoning mode gated to the $250-per-month Ultra tier, and API pricing near $1.25 per million input tokens and roughly $10 per million output tokens. Treat all of it as penciled in, not signed.
What's reportedly coming
The 2-million-token context window is the headline number, double the 1-million window on Google's current Gemini 3.5 Flash and, if it holds up, among the largest offered by any frontier model. Reported alongside it is Deep Think, Google's extended-reasoning mode, its answer to Anthropic's extended thinking and OpenAI's high-effort reasoning, where the model spends more compute on multi-step analysis before it answers. That mode is expected to sit behind the premium Ultra subscription rather than ship to every tier.
The model has also, by multiple accounts, had a rough road to the launch pad. It is roughly six weeks late, having slipped from a June target, and Google DeepMind reportedly scrapped an earlier base model after engineers found structural failures in recursive tool-calling and SVG generation. Independent trackers caution that even a genuine 2-million-token window does not guarantee the model reasons reliably across the full length, and early testers have reportedly found the Pro model trailing rivals such as Claude Fable 5 and OpenAI's GPT-5.6 on coding and long-horizon reasoning. Those are unverified claims until independent evaluations land.
The competitive timing is unforgiving. Gemini 3.5 Pro is expected to arrive a little more than a week after OpenAI shipped its GPT-5.6 family (Sol, Terra, and Luna) on July 9, and nine days after xAI's Grok 4.5, which reset the price floor at roughly $2 per million input and $6 output. Google's reported $1.25 input figure would undercut much of that field, a notable move in a year defined by the model layer cutting prices while the hardware layer compounds.
Why this matters
For Google, the stakes are strategic as much as technical. The launch lands amid a compute crunch severe enough that Google reportedly capped Meta's access to Gemini models this month after Meta asked for more capacity than Google could supply. That squeeze cuts both ways: it exposes how tight capacity is even for a hyperscaler, but it also underscores Google's structural edge. Google owns its models, its cloud, and its TPUs, so when capacity tightens its own roadmap gets first call on the chips. As Satvik Paramkusham, chief education officer at Build Fast with AI, put it, "the companies that own their compute will win the next two years. Everyone renting is one capacity crunch away from a stalled roadmap."
To matter against GPT-5.6 and the Claude line, Gemini 3.5 Pro will likely need to beat GPT-5.6 Sol on at least one headline benchmark, prove that its long-context recall actually holds at full length rather than degrading, and simply ship on schedule after a string of delay and talent-drain headlines. A cheaper price and a bigger context window are compelling on a slide; whether they convert to real capability is the open question independent testing will settle.
Then there is the geopolitics. The World AI Conference in Shanghai runs July 17 to 20 with more than 140 forums, over 1,100 exhibitors, and a heavy emphasis on global AI governance. Xi's decision to appear in person, after years of skipping the event, signals that Beijing now treats AI leadership as a top-tier national priority. Paired with ByteDance's new Seedream 5.0 Pro image model and Wall Street's growing embrace of Chinese models, the through-line of 2026 is a genuine two-superpower race. As Paramkusham framed the collision, "When a frontier launch and a head-of-state AI summit land on the same day, on opposite sides of the planet, that is the whole 2026 story in a single calendar square."
What to watch on July 17
Watch first for whether Google ships at all, or lets the date slip again. If it launches, the fast checks are whether an official model card confirms the 2-million-token window and the $1.25 input price, whether Deep Think is truly Ultra-only or carries a per-request surcharge, and how the model lands on independent long-context and coding benchmarks against GPT-5.6 and Claude Fable 5. And watch the split screen: a California model launch and a Shanghai governance summit sharing one square on the calendar. Whatever the benchmarks say, that image is the story of the year compressed into a day.
"When a frontier launch and a head-of-state AI summit land on the same day, on opposite sides of the planet, that is the whole 2026 story in a single calendar square."-- Satvik Paramkusham, Chief Education Officer, Build Fast with AI