Gemini 3.6 Flash makes the cheap tier good enough for agents

Google's new Gemini 3.6 Flash is the workhorse model, priced at $1.50 per million input tokens and $7.50 per million output. The improvements are concentrated where cheap models usually disappoint: agent and coding work. On DeepSWE it scores 49 percent against the previous Flash's 37, and on OSWorld-Verified it reaches 83.0 against 78.4. It also does the same work with fewer tokens, cutting output token use by about 17 percent, which lowers the real bill beyond the headline price. Computer use is now built in rather than bolted on.

Alongside it, Gemini 3.5 Flash-Lite targets high-volume agent loops at $0.30 input and $2.50 output per million, running around 350 output tokens per second. Its jump is sharper than the flagship's: Terminal-Bench 2.1 goes from 31 to 54 percent over the prior Flash-Lite. Google also shipped a third variant, 3.5 Flash Cyber, but kept it out of general release, offering it only to governments and trusted partners through a vulnerability-patching pilot. The pattern across all three is a deliberate split between a general model, a throughput model, and a restricted security model.

Why it matters

If you run agents at scale, the cost of a capable model per task keeps falling faster than the sticker price suggests, and Flash-Lite in particular makes long tool-use loops affordable. Before you keep routing everything to a frontier model, test whether 3.6 Flash clears your accuracy bar, because a 17 percent token cut compounds across millions of calls.