Gemini 3.1 Flash-Lite

“Best dollar-per-intelligence on Google's stack, now GA — the reservation is that Google already names 3.5 Flash-Lite as its replacement.”

For workloads that don't need frontier reasoning, 3.1 Flash-Lite delivers the best $/intelligence Google offers, with GPQA 86.9% proving it's not a toy. Vertex governance, VPC-SC, and Workspace integration are identical to the Pro tier, so there's no security trade-off. The strategic question has moved from preview status (GA since 2026-05-07) to lifecycle: Google names 3.5 Flash-Lite as its replacement and lists 2027-05-07 as the earliest shutdown date. Lock-in is the usual Google Cloud consideration — real but acceptable if you're already there. The 1M context lets it skip RAG in places Flash-Lite-class models historically couldn't.

“It quietly redraws the cheap-model frontier: reasoning-grade quality at extraction-grade price.”

In market terms, 3.1 Flash-Lite competes on the value frontier — the upper-left of the price/intelligence scatter — where GPQA 86.9% at $0.25 undercuts most rivals' quality-per-dollar. Its differentiation is offering reasoning capability and a 1M context at a price tier where competitors ship non-reasoning models with smaller windows. The moat, again, is Google distribution and Search grounding. Market timing is good as teams hunt for cost relief amid rising Flash prices. The weakness is positioning clarity: it overlaps 2.5 Flash-Lite on price below, and Google now names 3.5 Flash-Lite above it as its replacement.

“The cost winner in the lineup — $0.25/$1.50 with cache reads at 2.5 cents makes high-volume math trivial.”

This is the clear cost leader among reasoning-capable Gemini models — $0.25 input, $1.50 output, cache reads at $0.025. For high-volume pipelines it's roughly 6x cheaper than 3.5 Flash and ~12x cheaper than 3.1 Pro per token, and batch shaves another ~50%. The AI Studio free tier is generous for prototyping. Pricing is flat with no over-200K cliff, so it's predictable. GA on 2026-05-07 kept the preview prices unchanged; the finance caveat now is the 2027-05-07 earliest shutdown date, so budget a migration to 3.5 Flash-Lite, which costs more ($0.30/$2.50). Note audio input bills at 2x text ($0.50), which matters for transcription-heavy pipelines.

“Same SDK as the rest of Gemini, plus 332 tok/s — swapping it into a streaming pipeline is a one-liner.”

For builders, the appeal is zero migration cost (identical Vertex/AI Studio surface) plus genuinely fast output that changes streaming UX. Function calling and structured output via response schemas work reliably; the 1M context is preserved, which is unusual at this tier. The limits show up in longer agent loops, which feel less stable than on 3.5 Flash — keep tool chains short. For classification, extraction, and bulk transformation, ergonomics and speed are excellent.

“Fast and good enough for most things — I just don't reach for it on the hardest reasoning.”

End users rarely meet Flash-Lite directly; it lives inside apps and embedded assistants rather than as the Gemini app default. Where it surfaces, the perceptible quality is "good enough for most tasks" — fast, helpful, occasionally less thorough than the Pro tier on hard problems. The user-visible benefit is speed: responses stream noticeably faster than 3.5 Flash. Refusals follow the same Google policy stack as 3.1 Pro. For casual factual questions, Search grounding keeps it current. It's infrastructure more than a headline consumer model.

“An 86.9% GPQA at $0.25 is impressive — but it went GA with unpublished long-context recall and no SWE-bench on the card.”

The value story is real and the GPQA number is the genuine surprise. But the disclosure is thin: Google's launch blog gives GPQA and MMMU-Pro and little else — no SWE-bench, no AIME, no MRCR, no HLE — so coding and long-context claims rest on the family reputation, not this model's published evals. It went GA on 2026-05-07 at the preview price, and Google already names 3.5 Flash-Lite as its replacement. And the 1M context is advertised without any recall figure, which for a Flash-Lite model likely degrades earlier than the Pro tier. The honest read: an excellent cheap reasoning model whose data sheet is half-blank.