Where 3.6 Flash stands

For a regular user of Google Antigravity and hit the caps regularly, 3.6 Flash has been a big improvement. Past Gemini models have had the problem of not caring about token efficiency, which can be seen on DeepSWE with Gemini Flash 3.5 having around the same cost as GPT 5.6 Sol on Max reasoning, with the only reason it's close is due to lower token cost compared to OpenAI models.¹ On intelligence, it scores about the same as GPT 5.6 Terra on medium, which costs about a 10th as much.

Now, how has 3.6 Flash improved on 3.5 Flash? Let's start with token efficiency. In DeepSWE, it uses 97K output tokens per task compared to 277K for 3.5. While these numbers are a massive improvement, if you compare them to any other major model, the number are bad. The only model it's more efficient than is Fable 5 on max reasoning.

If we look at the other side of the equation, token cost, Gemini 3.6 Flash has a huge upgrade as every model upgrade normally has a cost increase. In a surprise, 3.5 Flash is more expensive per output token compared to its successor 3.6 Flash. For context, 3.5 Flash is $9 per million tokens of output compared to $7.5 per million tokens of output for 3.6 Flash.

Combining these two improvements, we get a cost reduction of about half for the same level of work, if not better.

Where's Gemini 3.5 Pro

With Gemini Flash 3.6 out before we even got Gemini 3.5 Pro, when and where is it going to land? With it originally going to be released June 2026, which was then pushed back to July 2026 and us only having a week left in July, is it going to get pushed back again?

While I can't say when it will drop, it's not going to be more than a month away hopefully. But as we have seen sub agents and workflows being used by Sol and Fable, is that the goal for Gemini 3.6? Gemini 3.6 Flash could have been refined from Gemini 3.5 Flash to use fewer tokens and cost less while working better as a sub agent. If my theory is correct, I could see Gemini 3.5 Pro releasing in the next week or two.

My prediction for 3.5 Pro is that it will land around GPT 5.6 Terra Max and Kimi K3 in performance while costing 1.2 to 2 times the price per task.

Conclusion

While Gemini 3.6 Flash is not a game changing model, it shows that Google is not fully falling behind. It could also be a sign of Gemini 3.5 Pro just being around the corner, with 3.6 Flash being designed to be a sub agent.