Gemini 3.5 Flash Review: Fast, Cheap, and Mostly Smart

Gemini 3.5 Flash is Google's new fast-and-cheap workhorse — $1.50/$9 per million tokens, a 1M-token context, and benchmark scores that punch well above its price tier. Google says it's "4x faster," and the streaming speed is real, but there's a latency catch nobody mentions. Here's the honest review: the benchmarks (and the one everyone confuses), the true speed picture, when it beats Gemini 3.1 Pro, and how to use it today.

At a glance

• What it is: Google's newest fast tier — the first in the Gemini 3.5 series. Released May 19, 2026.
• Price: $1.50 / $9 per million tokens ($0.15 cached) — cheaper than Gemini 3.1 Pro, yet ahead of it on several coding/agentic benchmarks.
• Smarts: 78.8% on the independent SWE-bench Verified board — strong for a "Flash" model.
• Speed: genuinely fast streaming (~153 tok/s), but a slow start (high time-to-first-token) at high thinking levels.
• Context: ~1M tokens in, 64K out, fully multimodal (text, image, audio, video, PDF).
• Not on the gateway yet — closest you can call today is Gemini 3.1 Flash / 3.1 Pro.

What's new

Gemini 3.5 Flash brings frontier-tier capability into Google's cheap, fast tier. The pitch: near-flagship intelligence at a fraction of the cost and several times the speed of previous Gemini versions, with strong agentic and coding behaviour. It keeps the hallmark Gemini features — a 1M-token context, full multimodality, and "thinking levels" that trade quality for cost/latency.

Pricing decoded

The pricing has a paradox the headlines love: it's more expensive than the previous Flash, but cheaper than Gemini 3.1 Pro — while often outscoring Pro on agentic work.

Benchmarks, honestly

Here's where most reviews trip up. There are two different SWE-bench numbers in circulation, and they're not the same test:

The 78.8% (SWE-bench Verified, independent) is the one to compare against other models — it places Flash fifth on the frontier board, an excellent result for a cheap, fast model. The 55.1% is Google's score on SWE-Bench Pro, a deliberately harder benchmark; quoting it next to other models' Verified scores (as some articles do) makes Flash look worse than it is. Different test, different scale.

Speed & the latency catch

Google markets Gemini 3.5 Flash as roughly 4× faster than other frontier models. The streaming throughput backs this up — Artificial Analysis measures around 153 tokens/second, genuinely quick. But there's a catch the marketing skips:

vs Gemini 3.1 Pro

How to use it today

Gemini 3.5 Flash is available directly via Google's API. On DataLLM Lab the current Gemini lineup is Gemini 3.1 Pro ($2/$12), 3.1 Flash, and 3.1 Flash Lite — the closest you can call through the gateway right now. When 3.5 Flash lands, it's a one-line swap:

from openai import OpenAI
client = OpenAI(base_url="https://www.datallmlab.com/v1", api_key="$DATALLMLAB_API_KEY")

resp = client.chat.completions.create(
    model="google/gemini-3.1-flash",   # swap to gemini-3.5-flash when available
    messages=[{"role": "user", "content": "Summarize this PDF and extract the tables..."}],
)
print(resp.choices[0].message.content)

For where Flash-class models fit among coders, see the best coding LLM guide; for the cheapest options overall, the cheapest LLM APIs.

FAQ