Google DeepMind is targeting July 17 for the general availability of Gemini 3.5 Pro, the rebuilt frontier model the company chose to construct from scratch after a wave of senior researcher departures and a high-profile June miss shook Alphabet’s market valuation. The launch lands on the same day DeepSeek plans to graduate its V4 family from preview to official stable release, setting up the most consequential week of the year for AI developers — and the practical deadline that matters most is July 24, when DeepSeek’s legacy chat and reasoner aliases stop responding with no announced extension.
The internal decision to abandon the Gemini 2.5 Pro base model entirely and run a completely new pre-training cycle is the headline most observers missed. Google reportedly could not close three performance gaps — mathematical reasoning, scalable vector graphics scene generation, and overall image quality — through incremental fine-tuning. Running a frontier-scale pre-training cycle from a clean base costs hundreds of millions of dollars and months of GPU time. Google chose that path anyway, which signals either how far short the prior candidate fell or how high the competitive bar has become.
Why the Rebuild Was Necessary
The timing compounds the pressure. At Google I/O on May 19, CEO Sundar Pichai told the audience to “give us until next month” for Gemini 3.5 Pro. The June window closed without a launch. The miss landed in the same two-week stretch that saw Noam Shazeer, the Gemini co-lead and co-author of the 2017 “Attention Is All You Need” paper that introduced the transformer architecture, announce his departure for OpenAI on June 18. John Jumper, the Nobel laureate behind AlphaFold and a nine-year DeepMind veteran, followed him out the door to Anthropic on June 19. Two additional senior researchers departed in the same window. Together, the exits triggered a five-percent single-session drop in Alphabet shares on June 22, erasing roughly $225 billion in market capitalization.
The rebuilt model is reported to feature a 2 million token context window — double the 1 million cap on Gemini 2.5 Pro — along with a Deep Think reasoning layer for multi-step logic and autonomous workflow capabilities for chaining complex coding and tool-use tasks. These specifications come from third-party reporting and leaks rather than official documentation. As of July 7, the public Gemini API lists only gemini-3.5-flash and gemini-3.1-pro-preview. No model card, pricing confirmation, or benchmark has been published. July 17 is a widely reported target, not an announcement with a signed launch post behind it.
What 2 Million Tokens Actually Means
A 2 million token context window lets a model hold roughly 1.5 million words in a single inference pass — a full large codebase, a year of meeting transcripts, or a multi-volume research dataset. Transformer attention scales quadratically with sequence length, so extending context to that scale requires significant architectural work. Researchers at Microsoft demonstrated a technique called LongRoPE that extends context to 2 million tokens, but reliable retrieval across the full span is a separate problem from technically accepting the input. Stanford researchers have documented a phenomenon where model performance degrades for information located in the middle 50% of a very long context, regardless of whether the model technically fits it.
- Effective context window often falls well short of the advertised limit.
- The evaluation to watch is whether reasoning quality holds across the full range, not whether the model accepts the prompt.
- Until independent evaluators run long-context retrieval benchmarks on Gemini 3.5 Pro, the headline is a capability claim rather than a verified specification.
The DeepSeek Factor
DeepSeek V4-Pro went live as a preview on April 24, 2026 — the same day OpenAI shipped GPT-5.5, a timing that appears deliberate. The model uses a Mixture-of-Experts architecture with 1.6 trillion total parameters, of which a routing network activates only 49 billion per token. That selective activation makes a 1.6-trillion-parameter model economically viable to serve, because each token costs roughly the same compute as a much smaller dense model while the full parameter count remains available across diverse tasks. V4-Flash ships with 284 billion total parameters and 13 billion active. Both variants run on Chinese-controlled infrastructure, which carries export-control and procurement implications that no contract clause overrides.
SpaceXAI’s Grok 4.5 sits in private beta with no confirmed public launch date. Canary strings naming the model appeared in the Grok web UI on July 6, and a subscriber rollout may follow at any point. The competitive calendar is now compressed: three of the most consequential frontier model events of 2026 land within a single week, with one more lurking behind a closed beta.
What Developers Should Watch
For any team running production systems on DeepSeek’s API, July 24 is the binding deadline. The legacy deepseek-chat and deepseek-reasoner aliases stop responding with no announced extension, which forces a migration to V4-stable or to an alternative provider. The migration window is roughly one week shorter than ideal, and the version that ships on July 17 may not have full third-party benchmark coverage by the time the legacy aliases go dark.
The right preparation move is to keep DeepSeek V4-Pro on a parallel evaluation track this week, regardless of whether Google ships Gemini 3.5 Pro on schedule. For teams that depend on long-context retrieval, treat any model card claim above 1 million tokens as a hypothesis to test rather than a specification to plan around until independent retrieval benchmarks exist. The frontier model calendar has accelerated to the point where any single launch is best understood as one input into a multi-vendor resilience strategy, and the next seven days will set the price, capability, and availability baseline for the rest of the year in AI infrastructure.