Intel / article

Gemini 3 6 Flash 3 5 Flash Lite Halving Time

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Jul 21, 2026

Context

Artificial Analysis tested two new Google models before launch and found that the shared headline, roughly halving the time per task, hides two different trade-offs. Gemini 3.6 Flash gets faster and a little cheaper without becoming more capable, while Gemini 3.5 Flash-Lite becomes much more capable but costs about twice as much per task, according to the Artificial Analysis write-up.

Gemini 3.6 Flash scores 50 on the Artificial Analysis Intelligence Index, the same as Gemini 3.5 Flash, with small moves in individual evaluations in both directions. Its average time per task fell from 2.7 minutes to 1.3, which Artificial Analysis attributes to the model spending fewer on each task and generating them faster, about 304 output tokens per second in pre-launch testing. Cost per task fell about 18 percent, from $0.59 to $0.50, because the model writes less output and Google cut the output price from $9.00 to $7.50 per million tokens while input stayed at $1.50.

Gemini 3.5 Flash-Lite replaces Gemini 3.1 Flash-Lite, and the jump is larger: its index score rose from 25 to 36, with the biggest gains on agentic evaluations such as TerminalBench v2.1, up 22.5 points. Average time per task fell from 1.0 minute to 0.6. The cost rise comes from pricing rather than usage. The model actually produced fewer output tokens per task, about 13,000 instead of 20,000, but Google raised its price from $0.25 and $1.50 to $0.30 and $2.50 per million input and output tokens, so average cost per task went from $0.04 to $0.09.

Both models keep a one-million-token and accept text, image, video and speech while answering only in text. The speed figures come from pre-launch testing, and the article does not describe which tasks make up its per-task averages.

Key takeaways · AI-distilled

Gemini 3.6 Flash scores 50 on the Intelligence Index, matching 3.5 Flash, with GDPval-AA v2 up 72 to 1421 and HLE down 3 points to 38%. Artificial Analysis credits its faster tasks to token efficiency and 304 output tokens per second in pre-launch tests.
Flash's cost per task fell about 18%, from $0.59 to $0.50, from lower output token use and new pricing of $1.50/$7.50 per million input/output tokens, down from $1.50/$9.00.
Gemini 3.5 Flash-Lite's 11-point gain to 36 came mostly from agentic : GDPval-AA v2 rose 498 to 1140 and TerminalBench v2.1 rose 22.5 points to 53.6. It averages 0.6 minutes per task at about 350 output tokens per second.
Flash-Lite's per-task cost rose from $0.04 to $0.09 even though output fell from 20k to 13k tokens per task, because pricing moved to $0.30/$2.50 per million from $0.25/$1.50. Both models keep a 1M context and a 90% cached-input discount.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source artificialanalysis.ai