Summary: Google introduced Gemini 3.1 Flash Lite in preview on March 3, 2026 as the fastest and least expensive Gemini 3 model, priced at $0.25 per million input tokens and $1.50 per million output tokens. The New Stack reports throughput up to 363 tokens per second. Google published no agent benchmarks, and no general-availability date was announced.
March 3, 2026: Google lifted the curtain on Gemini 3.1 Flash Lite in preview, reachable via the Gemini API inside AI Studio and on Vertex AI. The pitch leaves little ambiguity — across the Gemini 3 series, this is the model that runs fastest and costs least to operate, aimed squarely at developer workloads moving serious volume. The release pattern breaks with habit as well: the Gemini 3 generation shipped no Flash-Lite at all, and this cycle Google passed over a standard Flash to get the Lite variant out first, landing roughly two weeks after Gemini 3.1 Pro arrived on February 19. Read that sequencing as a statement about where Google expects demand to sit: cheaper inference tokens, sooner.
By VentureBeat’s reckoning, running it costs about one-eighth of what Pro does. For any product that meters charges by the token, this is more than an incremental tweak — the model-selection matrix gains a fresh cell, and one that plenty of high-volume pipelines will struggle to pass over.
Cost and Throughput
You pay $0.25 per million input tokens and $1.50 per million output tokens, the floor of the Gemini 3 family. The previous generation tells a different story: Gemini 2.5 Flash-Lite went for $0.10 and $0.40, so the newcomer bills several times what its predecessor did — that spread is the price of third-generation capability. Anyone budgeting a migration should run the numbers explicitly rather than assume “Lite” means cheap in absolute terms.
On throughput, The New Stack reports up to 363 tokens per second, quicker than any other Gemini 3 model. That sits three tokens per second under the old 2.5 Flash-Lite, yet still two to five times the pace of tier-matched rivals. Google’s own figures, credited to Artificial Analysis, claim time to first answer token cut by 2.5x and output speed raised 45% against 2.5 Flash, with quality holding steady or better. Where perceived latency in interactive products hinges on time-to-first-token, that pairing — cheaper than Pro, quicker than the field — is the actual product on sale.
Benchmark Scores and the Rivals
Flash Lite turns in an Elo of 1432 on the Arena.ai leaderboard. It posts 86.9% on GPQA Diamond and 76.8% on MMMU Pro, finishing ahead of same-tier models — and ahead of some larger prior-generation Gemini models — on reasoning and multimodal understanding alike. The New Stack’s side-by-side: it beats Gemini 2.5 Flash overall while costing less; GPT-5 mini and Claude 4.5 Haiku, its head-to-head rivals, land behind it; Grok 4.1 Fast undercuts it on price but trails badly on speed.
Worth flagging: Google published no agent benchmarks. With the rest of the market busy casting small models as agent workhorses, that silence reads as a signal about where Google believes this model belongs — and where it does not. Also worth knowing: the consumer Gemini app will not carry this model. It is purely an API product, which keeps the consumer experience and the developer cost story on separate tracks.
Reasoning Depth You Can Dial
Flash Lite ships with adjustable thinking levels: per task, developers set how hard the model reasons. Across high-frequency workloads this is the cost lever that matters most — drop the reasoning level, emit fewer tokens, and the bill shrinks to match. Reasoning depth stops being a fixed property of the model and becomes a per-request engineering call, the same lever OpenAI exposes through its reasoning-effort parameters.
Google’s demos span filling an e-commerce wireframe with hundreds of products, building a real-time weather dashboard from live and historical data, running a multi-step SaaS business agent, and rapidly analyzing and sorting large image sets. Latitude, Cartwheel, and Whering signed on early, and their feedback converges on efficiency and instruction adherence — the two qualities that count most when a model anchors the bottom of a high-volume pipeline.
Where It Fits
A workable split looks like this. High-volume data processing, translation, and content moderation, together with UI and dashboard generation and complex instruction following — that territory belongs to Flash Lite. For orchestration layers coordinating fleets of agents, Google itself offers no benchmark data; Pro or a purpose-built model is the sounder pick until someone publishes credible agent numbers. Teams currently running on 2.5 Flash-Lite should do the price math before migrating: the bill climbs several-fold, and what you buy is third-generation multimodal understanding and reasoning — a trade that pays off or not depending on how sharply your traffic mix reacts to speed and quality. Specs can still shift during preview, and no general-availability date is on the calendar, so pin versions in production and re-verify once the GA build lands.