Gemini 3.6 Put a 2M Context Window on Every Paid Tier. Here Is What It Actually Costs
Gemini 3.6 Pro shipped July 14 with the 2 million token context window available on every paid tier rather than behind an enterprise agreement. That is a genuine democratisation of a capability that was previously gated, and it deserves the positive coverage it got.
It also deserves an arithmetic check, because "you can use a 2M context window" and "you can afford to use a 2M context window" are different statements.
The Arithmetic
Gemini 3.6 Pro charges $2.00 per million input tokens. Fill the window completely and you are at $4.00 for the input alone, per request, before generating anything. Do that a thousand times a day and you are spending $4,000 daily, or roughly $1.4 million a year, on context you are probably not using well.
Run your own numbers on the annual cost calculator. The number surprises people.
Context Caching Changes the Equation
The feature that makes long context economically viable is not the window, it is the cache. Gemini 3.6 charges $0.20 per million for cached input, a tenth of the fresh rate. If you are querying the same corpus repeatedly, the first request pays full price and subsequent requests within the cache lifetime pay a tenth.
That turns the thousand-requests-a-day scenario from $4,000 into something closer to $400 plus cache storage. Which is a completely different business case.
The catch is that caching only helps when the prefix is stable. If every request has a different document, you pay full price every time. Cache-friendly architecture means putting your stable corpus first and your variable query last, which is a prompt-structure decision you have to make deliberately.
Flash Is Where the Volume Should Go
Gemini 3.6 Flash keeps a 1M window at $0.50 per million input and $3.00 output. It scores 87.6 on MMLU and 88.2 on HumanEval, which for high-volume production traffic is more than adequate.
The throughput per dollar on Flash is the best of any hosted model near the frontier. If you are running a RAG front end, a summarisation pipeline or a classification service, Flash is very likely the correct answer and Pro is very likely overkill.
Does the Window Actually Work?
Partially, and honestly. Retrieval accuracy across a full 2M window is good at the beginning and end and degrades in the middle, same as every long-context model. Google's needle-in-haystack numbers are strong. Real-world performance on documents where the relevant information is diffuse rather than a single findable fact is weaker.
Our standing advice has not changed: a large window is a convenience, not a replacement for retrieval. If you can find the relevant 20K tokens and pass only those, you will get better answers and pay less. Size your payloads first with the JSON token calculator.
The Genuinely New Capabilities
- Native video at 1 fps. Not frame extraction bolted on, actual video understanding. Best in class by a clear margin.
- Grounding with Search. Live web sources cited inline. Reliable, and the citation quality is good enough to audit.
- Thinking budget control. Explicit control over reasoning spend per request, which pairs well with cost-sensitive routing.
Compare Gemini 3.6 Pro and Flash directly on our comparison tool.