AI summary
Gemini 3.5 Flash prioritizes speed and agent capabilities, delivering stronger performance than earlier Gemini models and some competitors. The tradeoff is a roughly fivefold increase in running costs compared with Gemini 3 Flash.
The opening
Declan opens by teasing a tradeoff—Google’s focus on speed and agent capabilities has come at a cost—which sets up the video’s central tension.
Key points
- Artificial Analysis measured Gemini 3.5 Flash at about 280 output tokens per second in pre-release testing.
- The model’s speed puts it on the Pareto frontier of speed and intelligence, ahead of Gemini 3 Flash and GPT 5.4 mini in speed.
- On the GDPval agentic evaluation, Gemini 3.5 Flash recorded an ELO of around 1650, ahead of Gemini 3.1 Pro, Kimi K2.6, and GLM 5.1.
- Running Gemini 3.5 Flash costs around five times as much as running Gemini 3 Flash.
- The token price is three times higher: $1.50 per million input tokens and $9 per million output tokens; the model also uses more tokens and turns on agentic evaluations.
Quotes
We measured in our pre-release testing at about 280 output tokens per second.
Speed and intelligence is really improved, but cost is also really increased.
Caption and hashtags to share with
Gemini 3.5 Flash brings a major speed and agent-capability boost—but at around five times the running cost of Gemini 3 Flash.
- #gemini35flash
- #artificialanalysis
- #aiagents
- #llmevaluation
- #aiinference
- #generativeai
Written by gpt-6-luna from the transcript. Check the transcript for anything important.
Transcript
- Hey, my name is Declan Jackson. I am a member of technical staff here at Artificial Analysis, and I'm going to do a quick chat through the recent release of Gemini 3.5 Flash. This release is really interesting because Google have really prioritized two things with this release, but it has come at a bit of a cost. With this release of Gemini 3.5 Flash, Google have focused on speed and agent capabilities.
- In terms of speed, Gemini 3.5 Flash we measured in our pre-release testing at about 280 output tokens per second, which is pretty impressive for a model of that level of intelligence. This really puts it on the Pareto frontier of speed and intelligence. This also is a massive jump from Gemini 3 Flash speeds and puts it ahead of a model like GPT 5.4 mini.
- On agentic capabilities, it has a massive uplift from Gemini 3 Flash, and even more so than Gemini 3.1 Pro. Agentic capabilities has been a bit of a weakness for Google in the past. So it's really good to see that they've uplifted that, especially as we've seen in our real-world task agentic eval, GDPval: Gemini 3.5 Flash records an ELO of around 1650. This is ahead of Gemini 3.1 Pro and other models like Kimi K2.6 and GLM 5.1.
- But as I mentioned, this does come with a bit of a trade-off. The model costs around five times as much to run compared to Gemini 3 Flash. This is made up of two factors. First, the actual token price is a lot higher.
- The token price is three times that of Gemini 3 Flash, at $1.50 per million input and $9 per million output. We also find that it is using tokens on these evaluations. So it's reasoning more and it's also using more turns on our agentic evaluations, which is going to mean that it's going to cost more to run. So it's a really interesting trade-off here. Speed and intelligence is really improved, but cost is also really increased.
Clean adds punctuation and paragraphs and drops filler words. Switch to Original for the words exactly as transcribed.