Google's Best Reasoning Mode Is Landing Behind a $250-a-Month Subscription. Before You Budget for Deep Think, Here's What a Solo Operator Actually Needs.
Google announced Gemini 3.5 Pro at I/O on May 19 with a 2-million-token context window and a reasoning mode called Deep Think, told everyone to "give us until next month," and the headline that's settling into place is the price of the good part: Deep Think is gated to the $250-a-month Ultra subscription, not the $20 Pro plan. The frontier didn't just get more capable. It grew a velvet rope, and the most powerful reasoning sits on the far side of it.
Worth being precise about where this actually is, because the marketing is ahead of the rollout. As of mid-June, Gemini 3.5 Pro is still in limited preview for select Vertex AI enterprise customers. It has not shipped broadly to the consumer app or to everyone via the API. So this isn't a "go switch your stack today" post: the model isn't fully out. It's a "here's what the pricing structure tells you, and how to think before the bill is a real option" post. The structure is the story even before general availability lands.
What's behind the rope
Strip out the launch choreography and three facts matter. The context window is 2 million tokens, which is genuinely large and useful for whole-codebase or whole-corpus work. The API pricing is reported around $15 per million input tokens and $60 per million output tokens, roughly ten times the cost of Gemini 3.5 Flash. And Deep Think, the extended-reasoning mode that's the actual reason to care about Pro over Flash, is reserved for Ultra subscribers at $250 a month.
That last one is the design decision worth sitting with. Google is telling you, in pricing, that base Pro is the product and deep reasoning is a premium SKU. The $20 Pro plan gets you the model. The $250 plan gets you the model thinking hard. Reasoning depth is no longer a default you get for using the good model: it's a tier you buy.
This is a pattern, not a Google quirk
It would be easy to read this as Google being greedy. It's more useful to read it as the shape the whole market is taking. Extended reasoning is expensive to serve: it burns far more compute per response because the model generates and evaluates long internal chains before answering. Vendors across the board are responding the same way: meter it, tier it, or both. The names differ; the move is identical. The cheap tier gives you a fast answer. The expensive tier gives you a slow, deliberate one, and you pay for the deliberation.
For a solo operator, the trap isn't the price. The trap is assuming you need the expensive tier because it's the best one. "Best reasoning model" triggers the same reflex as "best laptop": you reach for the top of the line because it's the top of the line, not because your workload demands it. At $250 a month, that reflex is a $3,000-a-year line item, and most solo workloads will never touch the ceiling it's paying for.
How to tell if you're actually reasoning-bound
Here's the honest diagnostic, and it's not flattering to how most of us reach for these tools. The majority of tasks people throw at a frontier model and find disappointing aren't reasoning-bound. They're context-bound or scaffolding-bound. The model gave a weak answer not because it couldn't think hard enough, but because it didn't have the right information in front of it, or the task wasn't broken down, or the prompt asked for everything at once.
A genuinely reasoning-bound task looks like this: the inputs are complete, the question is well-specified, and the model still has to do real multi-step deduction to get there: a thorny refactor across an entire codebase, a proof, a planning problem with interacting constraints, an analysis where the steps genuinely depend on each other. Those exist, and Deep Think will earn its money on them. They are a smaller share of your week than the marketing implies.
The far more common failure is a task that a cheaper model would have nailed if you'd fed it the right context and split it into steps. Throwing Deep Think at a context problem just buys you a more expensive version of the same wrong answer. The fix there isn't a $250 subscription. It's better retrieval, tighter prompts, and decomposition: scaffolding, which is free and which you control.
What I'd actually do
Don't budget for Deep Think yet. The model isn't generally available, and the workloads that justify it are rarer than they feel. Instead, do the boring diagnostic first: for the tasks where your current model disappoints, figure out whether the bottleneck is reasoning or context. Log a few real failures and ask honestly which kind they are. My bet, from doing this on my own work, is that most land on the context-and-scaffolding side, and the cheaper tier plus better inputs closes the gap for a fraction of the cost.
When Pro and Deep Think do reach general availability, treat them like a specialist you call in, not a retainer you pay monthly. Route the 5% of genuinely hard, reasoning-dense tasks to the expensive tier per-call through the API, and keep the daily driver on the cheap, fast model. Paying $250 a month flat only makes sense if you're hitting deep reasoning constantly, and if you are, you'd know it because you'd already be frustrated by the ceiling on cheaper tools, not because a launch post told you the rope existed.
The honest counter-take: there's a real case that a flat subscription beats per-call API billing for a heavy user, because metered frontier reasoning can blow past $250 fast if your volume is high, and a predictable monthly cost is its own kind of sanity. That's true, and if you genuinely live in deep-reasoning workloads all day, Ultra may be the cheaper and saner option. But that's a conclusion you reach from your own usage data, after the model ships and you've measured it, not a default you adopt because Google built a premium tier and the number sounded like what serious people pay.
Author
Lukas
@lukcombinator