Name: Gemini 3.5 Flash-Lite
Author: Google DeepMind

Gemini 3.5 Flash-Lite is Google DeepMind's cheapest paid model, released July 21, 2026 with Gemini 3.6 Flash and a gated security variant, Gemini 3.5 Flash Cyber. Google's own copy calls it "the most cost-effective model in the class," and the pricing backs that up: $0.30 per million input tokens and $2.50 per million output tokens, undercutting both GPT-5.4 mini and Claude Haiku 4.5 on input price. The model card lists its architecture as built on Gemini 3.1 Flash-Lite, the preview-only budget model Google shipped in March 2026, rather than a fresh design.

TL;DR

Prices at $0.30/M input and $2.50/M output tokens, the cheapest paid tier in the Gemini 3 line
1M-token context window; OSWorld-Verified jumps to 74.0% and Terminal-Bench 2.1 to 54.0%, both roughly double Gemini 3.1 Flash-Lite's scores
Beats GPT-5.4 mini on computer use and long-context retrieval, loses to it on Terminal-Bench 2.1 and knowledge-work Elo

Google is positioning Flash-Lite as the layer beneath Gemini 3.6 Flash rather than a competitor to it: 3.6 Flash for complex agent work, 3.5 Flash-Lite for the high-volume, latency-sensitive jobs that make up most production traffic, agentic search, document processing, translation, classification. It's already rolling out inside Google Search's AI Mode and shipped with named early adopters Ashler, Palo Alto Networks, and Ramp, all cited for using it to scale agentic pipelines where per-call cost adds up fast.

The generational story is more complicated than a single launch date suggests. Google's own deprecation schedule still lists Gemini 3.1 Flash-Lite, not 3.5, as the named migration target for Gemini 2.5 Flash-Lite, which shuts down October 16, 2026. But 3.1 Flash-Lite never left preview status and was superseded just four months after it shipped, so any team migrating off 2.5 Flash-Lite today lands on 3.5 Flash-Lite as the current cheapest stable option in the Gemini 3 family.

Key Specifications

Specification	Details
Provider	Google DeepMind
Model Family	Gemini
Parameters	Not disclosed
Context Window	1,048,576 input tokens / 65,536 output tokens
Input Modalities	Text, image, audio, video
Output Modalities	Text
Knowledge Cutoff	March 2026 (some domains limited to January 2025)
Input Price	$0.30 / M tokens (standard)
Output Price	$2.50 / M tokens (standard)
Release Date	July 21, 2026
License	Proprietary

Benchmark Performance

Every number below comes from Google DeepMind's official Gemini 3.5 Flash-Lite model card, which publishes GPT-5.4 mini and Claude Haiku 4.5 scores from its own evaluation harness with the Gemini figures. Treat the cross-vendor columns as directional rather than neutral, the same caveat that applies to every vendor-published comparison.

Benchmark	Gemini 3.5 Flash-Lite	Gemini 3.1 Flash-Lite	GPT-5.4 mini	Claude Haiku 4.5
SWE-Bench Pro (Public)	54.2%	38.3%	54.4%	39.5%
Terminal-Bench 2.1	54.0%	31.0%	59.2%	44.2%
MLE-Bench	39.2%	22.0%	n/a	n/a
GDPVal-AA v2 (Elo)	1140	642	1171	907
OSWorld-Verified (computer use)	74.0%	54.3%	72.1%	50.7%
GDM-MRCR v2 (128K, avg)	72.2%	60.1%	42.7%	35.3%
GDM-MRCR v2 (1M, pointwise)	21.3%	12.3%	n/a	n/a

The jump over Gemini 3.1 Flash-Lite is the headline: SWE-Bench Pro nearly closes a 16-point gap in one generation, and OSWorld-Verified climbs almost 20 points. Against outside competitors the picture splits by task. GPT-5.4 mini still wins on Terminal-Bench 2.1 and GDPVal-AA v2, the benchmarks that reward sustained multi-step reasoning, and edges out Gemini narrowly on SWE-Bench Pro. Gemini 3.5 Flash-Lite pulls well ahead on OSWorld-Verified and by a wide margin on long-context retrieval, where GPT-5.4 mini's 128K score sits 30 points behind. For teams choosing a budget model purely on coding benchmarks, GPT-5.4 mini remains the stronger pick. For document-heavy or long-session agent work, the context numbers favor Gemini. Check the cost efficiency leaderboard for how this stacks up against other budget models as more independent runs come in.

Bar chart showing Gemini 3.5 Flash-Lite's gains over Gemini 3.1 Flash-Lite on Terminal-Bench 2.1 and GDPVal-AA v2 Google's own evaluation chart: Terminal-Bench 2.1 rises from 31.0% to 54.0%, and GDPVal-AA v2 Elo nearly doubles from 642 to 1140. Source: blog.google

Key Capabilities

Configurable Thinking for Subagent Work

Flash-Lite ships with three thinking levels, minimal, low, and higher, that developers set per request rather than per deployment. Minimal favors raw throughput for classification and routing; higher spends more compute on multi-step subagent tasks inside larger agent pipelines. That flexibility is the point of a "Lite" model in 2026: it's built to be the cheap worker a bigger orchestrator model dispatches narrow jobs to, not a standalone chat model competing on general intelligence.

Computer Use and Long Context

OSWorld-Verified, Google's benchmark for autonomous desktop and browser control, reached 74.0%, ahead of both GPT-5.4 mini and Claude Haiku 4.5 in the vendor's comparison. Paired with the 1M-token context window, that's the combination pushing Flash-Lite into agentic search and document-heavy workflows. GDM-MRCR v2 at 128K tokens hit 72.2%, more than double GPT-5.4 mini's 42.7%, though the 1M-token score of 21.3% shows the familiar drop-off every long-context model shows at the extreme end of its window.

"Gemini 3.5 Flash-Lite landed on the Pareto frontier in our receipt extraction benchmark, offering one of the best tradeoffs we've tested between accuracy, latency, and cost." - Veeral Patel, Head of Applied AI, Ramp

Multimodal Input at Volume

Text, image, audio, and video inputs run through the same endpoint at the same per-token price, which matters for the workloads Google is targeting: parsing scanned receipts, transcribing calls, reading charts embedded in reports. Palo Alto Networks' Ashwin Kannan, principal AI engineer, described the jump from Gemini 3.1 Flash-Lite as unlocking "a highly responsive option they can choose on demand for smooth, uninterrupted execution," language that lines up with the benchmark table: same price tier, meaningfully higher scores across almost every category.

Pricing and Availability

Standard pricing is $0.30 per million input tokens and $2.50 per million output tokens, confirmed against Google's Gemini API pricing page. Three other tiers apply depending on latency and reliability needs:

Batch and Flex: $0.15/M input, $1.25/M output, half the standard rate for asynchronous or best-effort jobs
Priority: $0.54/M input, $4.50/M output, for guaranteed low-latency serving
Context caching: $0.03/M tokens per cache hit on the standard tier, plus $1.00/M tokens per hour for cache storage

Against the competitors in the benchmark table, GPT-5.4 mini costs $0.75/M input and $4.50/M output, 2.5x Gemini's input price, while Claude Haiku 4.5 runs $1.00/M input and $5.00/M output, more than 3x Gemini's input rate. Gemini 3.5 Flash-Lite is the cheapest of the three on paper and it wins the benchmarks that matter most for high-volume, document-heavy pipelines, but GPT-5.4 mini's lead on Terminal-Bench 2.1 and GDPVal-AA v2 means the cheapest option isn't automatically the best one for agentic coding work.

The model is live now through the Gemini API, Google AI Studio, Android Studio, and the Gemini Enterprise Agent Platform, and is rolling out inside Google Search's AI Mode. No separate waitlist was reported at launch.

Strengths

Cheapest paid tier in the Gemini 3 lineup at $0.30/$2.50 per million tokens
Leads GPT-5.4 mini and Claude Haiku 4.5 on OSWorld-Verified computer use and GDM-MRCR long-context retrieval
1M-token context window with native text, image, audio, and video input
Configurable thinking levels let one model cover both throughput-first and reasoning-heavier subagent tasks
Broad day-one availability, already rolling out in Google Search AI Mode

Weaknesses

Trails GPT-5.4 mini on Terminal-Bench 2.1 (54.0% vs 59.2%) and GDPVal-AA v2 Elo (1140 vs 1171)
Long-context performance at the full 1M-token window drops to 21.3%, the same falloff pattern seen across the Gemini line
Google's own deprecation page still names Gemini 3.1 Flash-Lite, not this model, as the official migration target from Gemini 2.5 Flash-Lite
Parameter count undisclosed, standard for Gemini but limits architectural comparison
Text-only output, no image or audio generation

Gemini 3.6 Flash - the heavier sibling released the same day, for complex agentic and coding work
Gemini 2.5 Flash-Lite - the prior-generation budget tier retiring October 16, 2026
Gemini 3.1 Flash-Lite Review - our hands-on test of the preview-only model this one builds on
GPT-5.4 mini - OpenAI's closest price-tier competitor, ahead on coding benchmarks
Claude Haiku 4.5 - Anthropic's budget model, priced higher on both input and output
Cost Efficiency Leaderboard - live rankings of performance per dollar across budget and frontier models

FAQ

What is Gemini 3.5 Flash-Lite?

Google DeepMind's cheapest paid Gemini model, released July 21, 2026. It's built for high-volume, latency-sensitive tasks like document processing, translation, and agentic search rather than peak reasoning.

How much does Gemini 3.5 Flash-Lite cost?

$0.30 per million input tokens and $2.50 per million output tokens at standard pricing. Batch and Flex tiers drop to $0.15/$1.25, while Priority tier costs $0.54/$4.50 for guaranteed low-latency serving.

Is Gemini 3.5 Flash-Lite better than GPT-5.4 mini?

It depends on the task. GPT-5.4 mini scores higher on Terminal-Bench 2.1 (59.2% vs 54.0%) and GDPVal-AA v2 Elo (1171 vs 1140). Gemini 3.5 Flash-Lite leads on OSWorld-Verified computer use (74.0% vs 72.1%) and long-context retrieval, and costs less on both input and output tokens.

What is the context window for Gemini 3.5 Flash-Lite?

1,048,576 input tokens (about 1M) with a 65,536-token output limit. Retrieval accuracy is strong at 128K tokens (72.2%) but drops to 21.3% at the full 1M window.

Does Gemini 3.5 Flash-Lite replace Gemini 2.5 Flash-Lite?

In practice, yes, since Gemini 2.5 Flash-Lite retires October 16, 2026. Google's official deprecation page still names Gemini 3.1 Flash-Lite as the migration target, but that model was superseded by 3.5 Flash-Lite four months after its own launch.

Sources: