Gemini 3.7 Flash Is Here – And Speed, Not Intelligence, Is the Big Story

Google has released Gemini 3.7 Flash, and while it isn’t going to top any intelligence leaderboard, it’s one of the more interesting model releases of the year for a reason that has nothing to do with raw smarts: speed. Alongside it, DeepSeek has quietly shipped its first official coding harness, giving developers a browser-based alternative to tools like Codex and Claude Code. Here’s what both releases actually mean if you build with AI day to day.

What’s New in Gemini 3.7 Flash

Gemini 3.7 Flash arrives just weeks after Gemini 3.6 Flash and its cyber-focused variant, and it improves on that model while launching at an introductory price roughly 50% cheaper per million tokens than the original 3.6 Flash. That combination of “better and cheaper” is notable on its own, but the real headline is throughput.

On independent benchmarking platforms like Artificial Analysis, Gemini 3.7 Flash stands out as an outlier — running at around 340 tokens per second, dramatically faster than the pack. For comparison, even fast frontier models like GPT 5.6 Luna Max top out around 150 tokens per second, and top-tier reasoning models like Opus 5 Max run closer to 52 tokens per second. That puts Gemini 3.7 Flash somewhere in the range of six to seven times faster than a leading frontier model.

On pure intelligence, though, 3.7 Flash is not competitive with frontier-class models. It sits behind the GPT 5.6 “Data and Salt” class, behind models like Fable and Opus, behind Grok, and even behind open-weight models like Kimi and Qwen. Even Meta’s Muse Spark reportedly outperforms it on intelligence benchmarks. This is not a model built to win reasoning contests — it’s built to be fast.

Why Speed Actually Matters

It’s easy to dismiss a “fast but not that smart” model, but speed changes the shape of how you actually work with AI — and that’s the real case for getting excited about this release.

Developers have gotten used to AI being slow: kicking off a task, walking away, and coming back later. But for consumer-facing AI applications — chatbots, support systems, anything with a real person waiting on the other end — response latency matters enormously. A model that’s six to seven times faster doesn’t just save time; it changes the feedback loop. Faster responses mean more iterations per session, which means more actual work getting done in the same sitting.

The math plays out concretely: work that might take a slower frontier model six hours could take Gemini 3.7 Flash around an hour, plus maybe 30 minutes of additional prompting to correct mistakes — landing at under two hours total for a comparable result. The same logic applies to consumer-facing UX: a response that takes 14 seconds on a slower model could take roughly 2 seconds at a 7x speed boost, which is the difference between an app that feels sluggish and one that feels instant.

Putting It to the Test

In a real-world trial — asking the model to perform a security review of a codebase — the results were mixed at first. Code review tasks involve heavy tool-calling and file searching, which slows things down regardless of raw token speed, and the initial response felt slower than expected. But once the model started streaming in earnest, it completed roughly 90,000 tokens of work in about 2 minutes and 38 seconds — a task that would likely take a slower frontier model 10 to 15 minutes or more.

The takeaway: Gemini 3.7 Flash isn’t the model you reach for when a task demands maximum reasoning depth, but for high-iteration, latency-sensitive work, it’s genuinely fast in practice, not just on paper.

DeepSeek Ships Its Own Coding Harness

Separately, DeepSeek has released its first official coding harness — a genuine surprise, since developers have historically had to rely on third-party tools like OpenCode to work with DeepSeek models. What makes this harness distinctive is that it runs entirely inside a web browser rather than the terminal.

Running an npx command spins up a local URL with an interface that feels visually similar to Codex or a ChatGPT-style chat window. From there, you can mount a local folder as your working directory, select a model, and adjust the effort level for a given task.

A few features stand out:

• Live performance metrics — time to first token, tokens per second, and cache hit rate are all displayed in real time, so you can watch cache efficiency climb as a session progresses.
• Full trajectory visibility — a detailed breakdown of exactly what the model is doing, including the system prompt itself. In testing, the harness automatically pulled in an agents.md file as a system reminder, and the entire prompt structure was fully visible and inspectable.
• Multi-model support — the harness isn’t locked to DeepSeek’s own models. Through an OpenRouter integration, it’s possible to run other models — including Gemini 3.7 Flash — inside the same interface. In one test, switching to Gemini 3.7 Flash through this harness produced responses in the 200-230 tokens-per-second range.
• Agent presets and plugins — the harness supports configurable agent setups and an expanding library of plugins.

The interface leans toward power users rather than beginners — it’s considerably more transparent and configurable than simplified, browser-based coding tools aimed at non-technical users. For developers who want full visibility into what a model is doing and the flexibility to swap between models mid-workflow, though, it’s a genuinely capable option, even for someone who typically prefers CLI-based tools over UI-driven ones.

The Bigger Picture

Neither release is trying to claim the intelligence crown, and that’s precisely what makes them worth paying attention to. Gemini 3.7 Flash makes the case that raw speed is an underrated feature in its own right — one that reshapes how much real work gets done per session, especially for latency-sensitive consumer applications. And DeepSeek’s new harness shows that model providers are increasingly building their own polished tooling rather than leaving the harness layer entirely to third parties, giving developers more transparency and more flexibility in how they actually use these models.