Google and OpenAI Debut Super Fast AI Models—Gemini 3.7 Flash Is Out, But GPT-5.6 Sol Ultrafast Is Invite-Only

We are officially entering the era of the millisecond war. This week, Google and OpenAI both dropped new models that focus less on sheer intelligence and more on velocity. Google launched Gemini 3.7 Flash, while OpenAI countered with an invite-only model dubbed GPT-5.6 Sol Ultrafast. For founders and builders, this isn't just another incremental update to ignore. It marks a fundamental shift in how we build autonomous systems.

The End of the Latency Tax

For the last two years, building an AI agent was a trade-off. If you wanted the agent to actually think before it acted, you had to pay a latency tax. You would sit there watching a terminal window as the model crunched through a reasoning chain, taking five, ten, or even thirty seconds to come back with a structured response. That delay is a UX killer for anything interactive.

Google’s 3.7 Flash is designed to kill that delay. It is a multimodal model that supports what they are calling 'thinking' capabilities, but optimized for speed. It allows developers to toggle between standard responses and extended reasoning. If you are building a coding assistant or a customer service agent that needs to verify its own logic before hitting 'send,' this is the kind of architecture that makes those workflows feel instant rather than clunky.

OpenAI’s Gated Guardrails

On the other side of the fence, OpenAI is playing its classic game of scarcity. GPT-5.6 Sol Ultrafast claims to be the fastest reasoning model they have ever built, but unless you are in the inner circle or have a massive enterprise spend, you are probably looking at a waitlist. From what we can see, Sol is focused on the same goal: reducing the time-to-first-token while maintaining the 'O1' style of internal chain-of-thought processing.

As a founder, I find the invite-only approach frustrating but predictable. OpenAI is clearly testing the infrastructure load that comes with high-speed reasoning. Reasoning models are compute-heavy. When you try to make a heavy model run at light speed, you are basically trying to drag a tractor at 200 miles per hour. It’s expensive and prone to crashing servers. Google, having their own silicon and data centers, seems more comfortable letting the public in on day one.

What This Means for Agentic Workflows

If you are building in the AI space right now, you are likely moving away from simple chatbots and toward 'agents'—systems that take an objective and execute a series of steps to achieve it. The bottleneck for these agents hasn't been the LLM’s knowledge; it’s been the loop. An agent might need to make five API calls and three reasoning steps to complete a task. If each step takes 10 seconds, the whole process takes over a minute. That’s too slow for a human to stay engaged.

With Flash 3.7 and Sol Ultrafast, that loop might drop to sub-ten seconds. This changes the design space for what we can build. We are moving from 'asynchronous' AI to 'real-time' AI. Imagine a developer tool that fixes your code while you are still typing the next line, or a research agent that can scan 50 documents and give you a synthesized summary before you’ve finished your coffee.

Competitive Pricing and the Race to the Bottom

Google is pricing Gemini 3.7 Flash aggressively. This is the part that should make builders happy. We are seeing a race to the bottom on price-per-token for high-tier models. Google knows that if they make it cheap and fast enough, they can capture the developer market that is currently addicted to GPT-4o mini or Claude Haiku.

The strategy here is clear: hook the builders on the ecosystem by making the 'thinking' models as cheap as the 'dumb' models used to be. If you can get reasoning for the price of a standard prompt, there is no reason not to use it for every single interaction in your app.

A Healthy Dose of Skepticism

Despite the excitement, we have to stay grounded. 'Fast' doesn't always mean 'right.' In my experience, when you optimize a model for speed, you often see an increase in hallucinations or a decrease in the nuance of the logic. Google claims 3.7 Flash maintains high benchmarks, but benchmarks are rarely a reflection of real-world production edge cases.

We also need to consider the 'black box' problem. When a model reasons 'faster,' it often means it is skipping steps or using a compressed version of its logic. If you are building for high-stakes industries like fintech or healthcare, 'fast' is a secondary metric. Accuracy still reigns supreme. I’m skeptical that we can have O1-level reasoning at Flash-level speeds without losing some of the critical depth that makes reasoning models valuable in the first place.

Takeaway for Founders

If I am starting a project today, I am looking at Gemini 3.7 Flash as my primary playground for agents. The accessibility trumps OpenAI’s waitlist every time. You can’t build a business on a model you don't have access to. However, the real lesson here is that latency is no longer a valid excuse for poor AI experiences.

The tooling is finally catching up to the vision. We have the speed, we have the reasoning, and we have the low cost. Now it’s on us to actually build something that people want to use, rather than just another wrapper around a fast API. Spend less time worrying about the benchmarks and more time figuring out how a 500ms reasoning loop changes your product's value proposition.

• Google Gemini 3.7 Flash is live and optimized for low-latency agentic tasks.
• OpenAI GPT-5.6 Sol Ultrafast is faster but currently restricted to a waitlist.
• The cost of 'thinking' models is dropping, making them viable for every-day UI components.
• Speed is a feature, but it shouldn't replace rigorous testing for logic and accuracy.

Read the original at Decrypt →