Google scrapped the base architecture for Gemini 3.5 Pro and rebuilt it from scratch — a decision that cost hundreds of millions of dollars, slipped the launch to a July 17 target, and wiped $225 billion off Alphabet’s market cap in a single week. As of July 8, no model card has shipped, no pricing is confirmed, and no API documentation exists. July 17 is a widely-reported internal target, not a public commitment.
The question for developers is not whether the delay is embarrassing for Google. It’s what you should actually do about it.
Why Google Threw Away the Base Model #
This wasn’t a polish delay. Google identified three performance gaps in the original Gemini 3.5 Pro candidate that it decided could not be patched: mathematical reasoning, SVG scene generation, and image quality. GPT-5.6 Sol and Anthropic’s Fable 5 had pulled ahead on all three. Rather than ship a flagship that would visibly trail the competition at launch, Google ran a completely new pre-training cycle from scratch.
That choice is significant. Pre-training at frontier scale costs north of $100 million and takes months of GPU time. The fact that Google absorbed that cost to fix quality gaps signals something real: they believe a weaker flagship would cost more in enterprise credibility than the delay already has. Whether that bet pays off in nine days is another question.
The Confirmed Capabilities (and the Caveats) #
The rebuilt model is expected to ship with a 2 million token context window — double Gemini 3.5 Flash — plus a Deep Think Reasoning Layer and autonomous workflow capabilities. The 2M context window means a single API call can hold an entire large codebase, three years of team Slack history, or a full legal case file. Deep Think is Google’s name for extended inference-time compute: the model reasons through parallel hypothesis paths before responding, producing measurably more accurate results on hard multi-step problems.
None of this is official yet. Everything comes from enterprise previews, leaks, and third-party reporting. The only things Google has confirmed publicly: they’re rebuilding the model and targeting July. No model card, no benchmark results, no API documentation, no pricing table. Treat every specific number as provisional until the official Gemini API changelog goes live.
What to Prepare Before It Ships #
If Gemini 3.5 Pro ships on schedule, the API will follow the same breaking-change pattern already established by Gemini 3.5 Flash. Get ahead of it now:
Replace The parameter changed from an integer to a string enum (thinking_budget
withthinking_level
."minimal"
,"low"
,"medium"
,"high"
). The default shifted fromhigh
tomedium
— quality and cost will both change if you do a naive model-name swap.Remove Setting them to non-default values returns a 400 error on thinking models. Strip them from your generation config.temperature
,top_p
, andtop_k
.Add Required for tool-call workflows.id
andname
to allFunctionResponse
parts.Budget 30–50% more tokens for long agent loops. Thought preservation is on by default, and reasoning context carries forward across turns. Your token cost estimates from earlier Gemini versions are wrong for 3.5 Pro.
Deep Think requires a new thinkingConfig
object in your generation config with a thinkingBudget
parameter. The structure mirrors Anthropic’s extended thinking API, so if you’ve already implemented that pattern, the concepts transfer directly.
The 2M Context Math You Need to Do #
Gemini 3.5 Pro is expected to price at roughly $15/$60 per million input/output tokens — approximately 10x the cost of Gemini 3.5 Flash. Before you treat 2M context as an automatic upgrade, pressure-test whether you actually need it. Most teams using Gemini for RAG pipelines don’t. Gemini 3.5 Flash already offers 1M context, beats Gemini 3.1 Pro on coding and agentic benchmarks, and runs four times faster. The 2M window matters for specific workloads: monorepo-scale code review, full-document legal analysis, long-session agent memory. If that’s not your use case, the 10x cost premium buys you nothing.
One Hidden Trap: Computer Use #
Gemini 3.5 Pro does not support Computer Use, Image generation, or Audio generation. If you’re running browser-control agents or desktop-automation workflows via the Gemini Computer Use API, you cannot migrate to 3.5 Pro — you stay on gemini-3-flash-preview
until Google adds that support. This is underreported and worth checking before you finalize any migration plan.
What to Do Right Now #
Stop treating July 17 as a hard date. Google has already slipped this model twice in 2026. Set a calendar alert. If the model ships with an official model card on the 17th, you’ll have a week to evaluate before the DeepSeek hard deadline hits on July 24 — that one is not a target, it’s a cutoff. deepseek-chat
and deepseek-reasoner
stop responding after July 24 UTC with no announced extension. Gemini 3.5 Flash is generally available today and already excellent — it beats last year’s Gemini Pro on the benchmarks that matter for most developer workloads. Use it. When 3.5 Pro ships with documentation, you’ll know what you’re actually getting. Until then, building a roadmap dependency on an unconfirmed date is the mistake to avoid.