Why Gemini 3 Flash is the model OpenAI is afraid of

When Gemini 3 Pro Preview launched, Sam Altman called a Code Red at OpenAI. At the time, the Gemini 3 hyped looked a little overblown: Gemini 3 Pro, while showing flashes of brilliance, just isn't that good yet.

But yesterday, Google launched Gemini 3 Flash Preview and torched everyone.

You can see above the Power Ranking scores for Flash 3 compared to OpenAI and Haiku's lightweight offerings. And here's how Flash 3 stacks up against the heavyweights:

Flash 3 isn't quite beating the flagships, but it's clearly in the same ballpark. And how about if we look at speed and value as well?

Flash 3 is (barely) faster than all but opus, but more importantly it's 5x less expensive than the cheapest of the flagships.

Back to the lightweight competition

Flash 3's only weakness is that it's kind of slow--slower than the much larger model Opus 4.5. But what happens if we turn off thinking? How much difference does that make? We have the results for all of the lightweights:

Flash 3 with thinking disabled is smarter, faster, and cheaper than its closest competition Haiku 4.5, and while it's not quite as smart as GPT-5 mini (with default reasoning), it's about the same price and 8x faster. Meanwhile, GPT-5 mini gives up significantly more intelligence with reasoning dialed back to minimal (the lowest 5 mini can go).

(Haiku doesn't change much with reasoning enabled or disabled, since Anthropic trains its models to "think out loud" in the response body when thinking is off.)

So zooming in on the question of "what should your coding model be," there's four reasonable answers and here's how they stack up in a single graph:

Grok Code Fast 1 is an amazing value, and remains the coding model in Brokk's free tier. But Flash 3 non-thinking dominates the others to take the crown as the new default model in Brokk's paid tier. Try it out today!