Gemini 3.5 Flash-Lite

What is disclosed

Google's card for Gemini 3.5 Flash-Lite, published on 21 July 2026, points one step back for the architecture. The model "is based on Gemini 3.1 Flash-Lite".¹¹1Google DeepMind, Gemini 3.5 Flash-Lite model card, 21 July 2026. Architecture: "Gemini 3.5 Flash-Lite is based on Gemini 3.1 Flash-Lite." Knowledge cutoff March 2026, with some domains limited to January 2025. Benchmark table against Gemini 3.1 Flash-Lite: SWE-Bench Pro (Public) 54.2% and 38.3%; Terminal-bench 2.1, Terminus-2 harness, 54.0% and 31.0%; OSWorld-Verified 74.0% and 54.3%; GDM-MRCR v2 (8-needle), 128k average, 72.2% and 60.1%; 1M pointwise, 21.3% and 12.3%. Price rows per 1M tokens: $0.30 and $2.50 against $0.25 and $1.50. https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/ It went into the Gemini API as a stable release the same day.²²2Google, Gemini API model page for gemini-3.5-flash-lite (input limit 1,048,576 tokens, output limit 65,536, stable, computer use in preview, Live API not supported), and the release notes entry of 21 July 2026 ("a low-latency, highly cost-effective subagent option designed for high-volume automation"). https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite and https://ai.google.dev/gemini-api/docs/changelog

The Gemini 3.1 Flash-Lite card in turn names Gemini 3 Pro as its base. It adds that the model was trained on Google's TPUs.³³3Google DeepMind, Gemini 3.1 Flash-Lite model card, 3 March 2026: "Gemini 3.1 Flash-Lite is based on Gemini 3 Pro"; "trained using Google's Tensor Processing Units (TPUs)". https://deepmind.google/models/model-cards/gemini-3-1-flash-lite/

Only the Gemini 3 Pro card describes a sparse mixture of experts.⁴⁴4Google DeepMind, Gemini 3 Pro model card: "a sparse mixture-of-experts (MoE)" transformer, with no parameter or expert counts. https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf Neither Flash-Lite card uses the term. There is no parameter count on either.

Knowledge runs to March 2026, though the card warns that some domains stop at January 2025.¹¹ The context window is 1,048,576 tokens, and a reply can be up to 65,536 tokens long.²²

What is new

Google set it against Gemini 3.1 Flash-Lite, scoring both through the API at high thinking, on the first attempt.¹¹⁵⁵5Google DeepMind, Gemini 3.5 Flash-Lite model evaluation, July 2026: Gemini scores are pass@1, run through the Gemini API for gemini-3.5-flash-lite "with high thinking and default sampling settings"; SWE-Bench Pro with "an internal version of the Antigravity harness"; Terminal-Bench 2.1 with the default agent harness, Terminus 2; OSWorld-Verified "averaged over 5 runs with a single attempt per run". https://storage.googleapis.com/deepmind-media/gemini/gemini_3-5_flash-lite_model_evaluation.pdf

The biggest jump is on Terminal-bench 2.1, from 31.0% to 54.0% under the Terminus 2 harness.¹¹⁵⁵ SWE-Bench Pro went from 38.3% to 54.2% on the public set. That run used an internal version of Google's Antigravity harness.¹¹⁵⁵ OSWorld-Verified, a computer use test, was averaged over five runs and reached 74.0%, against 54.3%.¹¹⁵⁵

At 128k tokens, MRCR v2 with eight needles averages 72.2%, up from 60.1%.¹¹ At the full 1M the score is 21.3%. The older model managed 12.3%.¹¹

The price went up. The card lists Gemini 3.1 Flash-Lite at $0.25 per million input tokens and $1.50 per million output tokens.¹¹

Using it

The id is gemini-3.5-flash-lite.²² The API page describes it as "optimized for high-throughput, low-cost execution for subagent tasks and document parsing".²² In the release notes it is a subagent option "designed for high-volume automation".²²

Input costs $0.30 per million tokens, at one rate for every type of input.⁶⁶6Google, Gemini API pricing, paid tier, standard rates per 1M tokens, page last updated 24 September 2026, read on 29 September 2026. The input row reads "$0.30 (text / image / video / audio)". https://ai.google.dev/gemini-api/docs/pricing Output costs $2.50 per million, thinking tokens included.⁶⁶

Cached input is $0.03 per million tokens, plus $1.00 per million tokens for each hour the cache stays stored.⁶⁶

The Live API is not supported. Computer use is, as a preview.²²