⚡ Gemini 3.7 Flash is Blazing FAST. But Here're a Few Things You MUST Know Before Using It

Introduction

GPT 5.6 Sol and Claude Opus 5/Fable 5 keep dropping new models like they’re fighting in a boxing ring, while Kimi K3 and Qwen3.8-Max jump in now.

Gemini has been standing outside looking like that friend who says, “I just haven’t fought yet,” and of course, nobody believed him.

Now Google has finally sent Gemini 3.7 Flash into the ring, and its first punch is hard to ignore: around 340 tokens per second, with a launch price much lower than Gemini 3.6 Flash.

You can also run Gemini 3.7 inside DeepSeek Harness and watch its speed, cache, and agent trajectory in real time.

I’ll show you how fast Gemini 3.7 really is, what DeepSeek Harness can do, and whether Google finally landed a punch strong enough to make everyone look back.

I. What Changed From 3.6 Flash

First, Gemini 3.7 Flash is not a new base model.

3.7 is built on top of 3.6 Flash, it's an algorithmic reasoning and execution upgrade. That's actually interesting, because the gains it made are bigger than you'd expect from a three-week iteration.

1. Updated Benchmarks

On coding, the improvements are real and measurable across multiple benchmarks:

The gap gets bigger on longer tasks. On DeepSWE v1.1, which covers longer software engineering sessions, 3.7 Flash jumped 16 points over its predecessor. That's where the real upgrade shows up.

On Code Arena, Gemini 3.7 Flash scored 1,588, ahead of Claude Sonnet 5 (1,541) and GPT-5.6 Terra (1,523) in Google's comparison.

2. Pricing Story With the Asterisk

The launch headline says "half the price of Gemini 3.6 Flash." That's technically true, but there are 2 things I need to say.

First: the discount expires. 

The $0.75/$3.75 rate is introductory pricing through December 31, 2026. On January 1, 2027, pricing rises to $1.50 per million input tokens and $7.50 per million output tokens, exactly what Gemini 3.6 Flash cost at launch.

You have roughly 4 months to evaluate whether the capability gains justify building production agents on this before the economics change.

→ Gemini 3.7 Flash is not the cheapest option in its performance tier. GPT-5.6 Luna is listed at $0.20/$1.20 per million tokens, and DeepSeek V4 Flash is $0.14/$0.28 → both well under the Google rate even at the promotional price.

II. How Fast Is Gemini 3.7 Flash in Real Use?

~340 tokens per second sounds impressive. And it is. But that number needs context before you build your expectations around it.

Key points

• Around 340 tokens/second in testing, with an average task time of 1.7 minutes.

• A real security review finished in about 2 minutes 38 seconds despite file reading and tool calls.

• One user test used roughly 20% fewer tokens than Gemini 3.6 Flash.

1. What "340 Tokens/Second" Means

In practice, Artificial Analysis measured a median time to first token of ~9.83 seconds for Gemini 3.7 Flash at high reasoning. So: you wait about 10 seconds before you see anything, then output arrives fast once it starts.

That speed figure measures decoding after the first token. It is not time to first token, end-to-end latency, or a p95 service guarantee.

For agent workflows where the model is executing tool calls and re-reading context on every step, this matters.

Here's the full picture:

2. What Happens When Tool Calls Are Involved

Benchmarks usually happen in a cleaner setup, while coding agents have more work to do. In a test by Mehul Mohan, Gemini 3.7 handled a real project security review with many files and tool calls.

It needed time at the start to read the project and invoke tools. Even with all of that extra work, the run finished in around 2 minutes 38 seconds across roughly 90,000 tokens.

The model still has to complete the steps between them. But once generation starts, the high decoding speed moves the whole workflow faster than you'd see with a slower model.

3. Does It Use Fewer Tokens Too?

In Stevibe's comparison between 3.7 and 3.6 Flash, Gemini 3.7 used around 20% fewer tokens on the same task.

This was a personal test, so you shouldn’t treat the result as a benchmark for every workload. Still, it’s interesting because Gemini 3.7 may run faster while also using fewer tokens in some tasks.

III. Understanding DeepSeek Harness. Why Is It Here?

Being fast is one thing. You still need somewhere flexible to turn that speed into a real workflow. That's why I introduce DeepSeek Harness there.

1. What DeepSeek Harness Is

DeepSeek released DeepSeek Harness v0.1 on August 13, 2026, the same day as Gemini 3.7 Flash, it’s an open-source runtime for assembling agents from replaceable models, tools, skills, sessions, sandboxes, and interfaces.

The core design idea is "everything is a plugin." Models, tools, session state, filesystems, sandboxes, the agent loop, orchestration, and even the UI are all swappable Cordis services.

You can change the model, change how the loop works, change how sessions are stored, all from config.

It's MIT-licensed, open-sourced on GitHub at github.com/deepseek-ai/deepseek-harness, and built by a team that only formed in May 2026.

2. What You Can Do With DeepSeek Harness Now

You'll need Node.js 22.15 or newer, then you run npx @deepseek-ai/dsh web and the browser interface opens at http://127.0.0.1:3080. Add an API key, pick a workspace folder, and you're in. From there you can:

• Choose the workspace where the agent works, it reads your project files as context

• Pick between four modes, each loading a different default plugin set:

  • Standard: the full coding agent, with file editing, shell, file and web search, skills, planning, goals, subagents, and workflows.

  • Code: everything in Standard plus the Code Mode SDK, so the model writes a short TypeScript program to chain multi-step operations together.

  • Minimal: just two tools, a persistent bash shell and a file editor, which is handy for clean model benchmarking.

  • Creator: built for authoring your own presets, with runtime inspection and in-memory plugin experiments on top of Standard.

• Switch models, plugins, agent presets, and workspace without rebuilding anything

• Watch the agent work in real time: tokens per second, time to first token, cache hit rate, and tool call trajectory, all visible during the task

IV. Running Gemini 3.7 inside DeepSeek Harness

This is where the two launches connect. DeepSeek Harness supports multiple model providers, so you can bring Gemini 3.7 Flash into the same environment and watch its real behavior.

Now I’ll bring Gemini 3.7 into this workspace and test it with a coding task that is heavy enough to make the speed more meaningful.

1. Connect Gemini 3.7 Through OpenRouter

Inside DeepSeek Harness, go to Settings → Models → Add Provider, then choose OpenRouter.

Enter your OpenRouter API key, click Apply, and when you open the model selector, Gemini 3.7 Flash will appear in the list.

From there, Gemini 3.7 runs inside DeepSeek Harness while you watch tokens per second, time to first token, cache hit rate, and trajectory, all in the same interface.

2. Rubik's Cube Test: What Real Speed Looks Like

Stevibe tested Gemini 3.7 Flash by asking it to build a 3D Rubik's Cube inside DeepSeek Harness. The task:

• Finished in ~27 minutes

• Ran through 52 steps

• Ran at about 120 tokens per second

So, that's lower than the 340 tokens/sec benchmark. And that's expected. The benchmark measures a model streaming one long answer. Real speed in complex workflows is lower than peak decoding speed.

But here, you can see exactly where time went. DeepSeek Harness's trajectory view shows every tool call, every context read, every step, so you can understand the shape of the workflow.

3. Mario Game Test

Another user, KisekiyaCodes, also tested Gemini 3.7 Flash after seeing many model comparisons online. This time, the task was to build a Mario game.

Gemini 3.7 was handling an interactive, multi-component project rather than a short piece of code.

That's the kind of task where first-pass quality matters more than raw speed, and where the coding benchmark gains start to show up in ways you can actually feel.

V. Is Gemini 3.7 Flash Worth Using?

Hmm, to be honest, it depends on what you're doing and how much you care about what happens after December 31, 2026.

When It Makes Sense

When to Be Careful

• If you're planning post-2026 production costs: the price doubles on January 1, 2027, so budget against the 2027 rate.

• If you're in the EEA, UK, Switzerland, or Nigeria: consumer access via Gemini Spark isn't available in those regions.

• If raw cost is the priority: GPT-5.6 Luna ($0.20 / $1.20) and DeepSeek V4 Flash ($0.14 / $0.28) both come in well below Google's rate, even at the promo price.

• If you need the highest reasoning accuracy: test Gemini 3.7 Flash against a stronger frontier model first. Speed won't cover for accuracy on tasks where mistakes are expensive.

Conclusion

Gemini 3.7 shows that model competition is starting to change. Benchmark scores still matter, but real use also depends on how fast a model responds and how well it keeps moving through a longer workflow.

DeepSeek Harness makes that easier to see because you can watch what happens during the task instead of only waiting for the final answer.

Gemini 3.7 may not win every test, but Google has put Flash in a much harder position to ignore. If this direction continues, the models people use most may simply be the ones that make them wait less while still doing the job well.

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

• FREE Claude Prompt That AUTO Turns Any Idea Into a VOX-Style Video (Full Guide)

• How to Use Gemini + NotebookLM to Keep Your AI Accurate Every Time (For Free)

• "Entry-Level" is Lying to You. Here's What the AI Job Market Actually Looks Like Now*

• Coaxing Unique AI Designs is A Growing Challenge. But This Viral Workflow Solved It*

• 7 Jobs Claude Code Does Better than ChatGPT at Work (And None of Them Are Coding)*

  *indicates a premium content, if any