Greetings earthlings,

Welcome to Spark Intelligence – your AI navigator in the creative and marketing world. Today we are dedicating the entire newsletter to a big subject - Reasoning Models. So for this deeper dive Jules has joined Emma to write this edition with his Chief AI Guru hat on. Let us know what you think of this format - poll at the bottom!

Let’s dive in.

Reasoning Models: What are they and why should you care?

If you’ve been using ChatGPT lately, you may have noticed it prompting you to use “reasoning.” For those on a paid plan (which we highly recommend), you now have a host of model options—like o1 and o3-mini—both of which boast “advanced reasoning.” And it’s not just OpenAI making moves: in the last few weeks, the AI world has seen the launches of Gemini 2.0, Claude 3.7, Grok 3, and DeepSeek R1, all billed as “reasoning models” too.

So, what sets these new models apart? And why should you care? And what do they mean for your workflow?

The shift toward reasoning

Older AI models aimed to give you lightning-fast answers, often leaning more on “intuition” of the most likely answer than logic. But “reasoning models” take a step back. Rather than instantly generating a response, they break problems down into logical steps. You might have to wait an extra 30 seconds or so for your answer—but it’s often deeper, more structured, and more accurate. We’ve seen these models excel at tasks like workshop planning, strategic problem-solving, and complex coding projects.

Prompting Evolves

Whereas earlier approaches encouraged detailed, step-by-step instructions (often called “chain-of-thought” prompts), these new models seem to respond best to a clearly stated goal with minimal constraints. The model then figures out the chain-of-thought on its own.

Greg Brockman, co-founder of OpenAI, recently shared how best to interact with these new reasoning models. Instead of piling on every detail in your instruction, Brockman suggests providing a clear, simple goal and desired output format. The model then “decides” how best to break down your request. This approach can make your interactions more productive—and the AI’s responses more flexible.

Where can I find a reasoning model?

OpenAI started the reasoning model arms race with the release of OpenAI o1 in December. Since then we’ve seen them release o3-mini (where do they get these names from?!) and ChatGPT 5 is thought to be only a couple of months away from release. Just go to the drop down in the top right of your ChatGPT window.

Google have not been wasting any time either, and their latest reasoning model Gemini 2.5 Pro Experimental topping many of the “evals” - that’s evaluation benchmarks to you and me, how people are measuring the performance of these models. How well these ‘evals’ translate into real world business tasks is a matter for another newsletter!

Finally, Anthropic’s latest model also incorporates reasoning, Claude 3.7. Unlike OpenAI and Google’s approach you don’t need to explicitly chose which model you want to use to get reasoning capabilities. They’re introduced a concept they are calling “hybrid reasoning”. More on that below:

What exactly Is “Hybrid Reasoning”?

“Hybrid reasoning” means the AI can respond quickly to simple prompts while also having the capability to slow down, break down complex tasks, and provide a highly structured solution when needed. It adapts on the fly, so you don’t have to switch between different models for different types of tasks.

OpenAI’s approach right now separates models: GPT-4o for quick answers and o1 for deeper reasoning. Anthropic’s Claude 3.7 attempts to unify those features, answering simple queries almost instantly while shifting seamlessly into deep analysis mode if the prompt calls for it. Sam Altman recently announced that once ChatGPT 5 is launched they will be moving the same way.

Imagine having an assistant that can pivot instantly: quick brainstorming for creative sessions, then thorough step-by-step reasoning for project planning. That’s the essence of hybrid reasoning—and it’s transforming how we interact with AI.

Implications for the future of Large Language Models

Claude 3.7 Sonnet signals a shift toward a more flexible, user-centric AI experience. Rather than forcing users to pick “fast” versus “logical,” the model itself decides how much time and computational effort to spend on each task.

• OpenAI has promised a similar solution with ChatGPT 5.0, unifying quick and complex responses into a single model.

• Google continues to refine Gemini, betting that multi-modal reasoning will be the next big step in AI capabilities. 2.0 is designed to handle text, images, and more, offering an “all-in-one” approach that Google hopes will capture enterprise markets.

• xAI’s Grok 3 shows how powerful these models can be without guardrails—and reminds us of the ethical and regulatory challenges on the horizon.

Will legislative bodies in the UK and EU push stricter regulations if models like Grok 3 can be easily misused? Or will the arms race lead AI developers to accelerate releases? Time will tell.

Our take: Why this matters

We talk to a lot of agency leaders and CMOs who feel like they’ve ticked the AI box - a few licences, a workshop, maybe even a prompt library.

But the real shift isn't in tools. It’s in thinking.

Reasoning models move the AI shift along further. They open up new ways to structure thought, not just generate copy.

For CMOs: Use them to reframe how your team builds strategy, questions assumptions, and collaborates across silos. This is a business change programme that’s a people challenge, not a digital challenge.

For agencies: This is your moment to move the conversation from productivity to transformation. Help your team learn how to think with AI — not just prompt it. Elevate the questions you ask, and the value you bring.

In other news - our immensely talented AI Coach Matthew Maxwell has been hitting the headlines

Matthew is going to exhibiting Something Rich and Strange, a “psycho-geographic exploration” of Shakespeare’s work in Shoreditch later this month. His work introduces the imaging power of generative AI to Shakespeare’s classic works.

Matthew teaches the Design and Strategy modules of our AI Accelerator. He is an award-winning creative director and artist - studied Fine Art at Oxford University, and Human Computer Interaction at Cambridge University. Right now he is completing a PHD in creative AI at Middlesex University, and this exhibition forms part of the research. Previously, he spent several years as a production designer in the film and TV industry in Los Angeles before returning to London to join the burgeoning interactive industry as a creative director. His award-winning work includes a Cannes Gold Cyber Lion and a BAFTA nomination. 

Matthew showed us behind the scenes on creative process with AI earlier this year - catch up here, it’s a delight.

Cheers,
Emma & Jules
Co-Founders of Spark AI

AI isn’t just tools - it’s a business shift. Spark AI helps agency leaders and CMOs lead strategically, not reactively, with confident teams and a clear roadmap for the future.

Learn more 👉 Explore AI for Marketers | Explore AI for Agencies