Why Betting On One AI Model Is A Leadership Mistake

This week, OpenAI launched ChatGPT 5.5, hot on the heels of the recent Anthropic launch, Opus 4.7. The pace of development in AI remains dizzying: recent developments from just the past few weeks from the major companies include not just these new model releases but Claude in Word, Claude in Copilot and Copilot “agent mode” for Word, Excel, and PowerPoint. This partnership with Microsoft rivals the desktop agent that Google have been steadily pushing with Gemini.

At the same time there have been really impressive improvements in the models coming out of China - DeepSeek V4, Kimi K2.6, Qwen 3.6. Whilst these are not quite the same as the models available from OpenAI and Anthropic, the gap is closing fast. This matters, because not every company needs to be using ‘frontier’ models (ones that operate on the very edge of what is technically possible). DeepSeekV4 and Kimi 2.6 deliver almost the same performance on a fraction of the cost.

A good way of thinking about AI is the models themselves and then the interfaces to get things done. GPT 5.5 and Opus 4.7 (and DeepSeekV4, Kimi 2.6 etc.) are the LLMs.
ChatGPT and Claude are the interfaces. Recent commercial deals such as the one struck with Microsoft have opened Word or Excel and Copilot to become the place where the interface with an LLM occurs.

How Should Leaders Be Thinking About This?

The incremental improvement for each new release shows that there is still some way to go with AI model improvements. Having your team go "all in" on a single model and not encouraging them to try the newly released models will put you at a disadvantage. A challenge is that once a team has adjusted to the license you provided for Claude, incorporating a new model into their practices takes time and may be difficult.

The new challenge on the horizon for leaders will be the token usage cost (more on that in a moment...), and one solution may be looking to other high-performing models at a lower cost. Having a good foundation for understanding the process of model selection and application will help leaders both avoid wasted investments and enable their teams to get the most out of the models they use.

The critical need this highlights is that your teams need to be set up to be ‘learning teams’. If they are not learning at least as fast as the pace at which the world around them is moving, they are on the path to extinction. Competitive advantage comes from creating an environment and space to learn so that they at the very least keep up with developments.

The Economics of AI Inside Organisations

AI uses tokens, and tokens come at a cost. Balancing a strategy that encourages experimentation with the commercial realities of the cost of using AI is the frontier facing many leaders this year.
The initial cost of AI (outside of the change management focus required to make your organisation successful) can appear to be licences for your teams, but this is just the starting point. In addition to this, for more complicated work it is necessary to purchase additional processing power.

Tokens are the small chunks of text that AI models read and write. Models like ChatGPT and Claude process and price their outputs based on how many tokens are used - the more complex the work, the more tokens are used. The higher-priced subscriptions include a higher token allocation but this can be exceeded pretty quickly. The operator then has the decision whether to wait four hours for their allocation to refresh or to pay to use tokens to complete a task in the meantime.

We are starting to see job interviews where negotiations are not just about salary and benefits but also about what access to tokens an engineer will have. You can be the brightest developer but without access to processing power, you will struggle to maximise your work.

Some organisations (particularly the large tech organisations) are "forward investing" by giving each of their staff large allocations of tokens each. There are stories of these companies using the metric of the volume of token usage being used to measure the individuals’ performance. It doesn't take a genius to work out that this measure of performance is hardly the correct one and will likely lead to a good case study on the principles of Goodhart’s Law.

The organisations that will succeed in the coming years are not the ones that deploy the most AI but the ones that master the economics of token usage.

^{Takeaways For Leaders}

Understanding the economics of token cost, usage, and performance is a skill leaders need. Firstly, understanding the underlying economics of human versus machine cost: if an employee is using an average of $200 worth of tokens per day or $73,000 per year, are these tokens delivering as much as an equivalently paid employee would deliver? Secondly, are tokens being spent effectively? Ensuring your teams understand and are using tokens efficiently matters.

Selecting a more powerful model that is necessary for a task burns through tokens. Ruben Hassid has an excellent Substack and has recently posted about how to stop burning through tokens in Claude. Tips include:
-Converting files before uploading (a PDF page costs up to 3,000 tokens to read). Instead, pasting the text into a Google Doc, downloading it as a .md file is 90% cheaper.
-Do your planning in Chat, and building in Cowork. Chat is cheap and Cowork is expensive.
-Use the phrase "ask me questions" instead of a long prompt. Keeping the prompt to 30 words max can lead to big savings
-Edit your message, don’t send a follow-up. Claude re-reads all of it. The Edit button replaces the old message. Cleaner and cheaper.
-Selecting the right model for each task. A quick question should be a chat with Haiku. Writing a report based on files is a Cowork task with Opus and building a chart from data should be Claude Code with Sonnet.

Link to Ruben’s ‘How To AI’ can be found here

AI Can Now Be You - The Question is Whether It Should

Mark Zuckerberg has been building an AI clone of himself that his staff can interact with. Training an AI clone on the way that you think and are likely to respond is not new. AI clones have been used by creators to amplify output (Mr. Beast), to engage fans (Paris Hilton), in advertising (David Beckham) and even as a paid ‘virtual girlfriend’ (Caryn Marjorie).

In all cases the intention is to be in more than one place at the same time, with the theory that if trained on enough data the AI will respond the way that you would. The image and video generation tool, RunwayML, has just released a tool that creates a clone from an image that it turns into video, which you can then send into a meeting on your behalf (!).

Over the past couple of years as AI note-takers became popular, some people believed (and possibly still do) that it's okay not to show up to a meeting and to send their note-taker instead. This turned out to be a somewhat controversial approach that many (human) meeting participants took a dim view of.
AI avatars attending meetings are likely destined to follow the same path. But other work with an AI avatar may be okay. For instance if you use a GPT as a thought partner, having a face with which you chat rather than just a voice as you work things out may be a preference.

^{Testing it Out}

I tried out creating an avatar of myself and then chatted with it in a video meeting. As I've not trained it on my voice, it was a little strange at first, but for me it felt better to chat with an ‘AI person’ than to chat with an audio AI.

It seems unlikely that this would be a trusted method for a leader to engage with his teams in the near future, but the potential application of a leader of a large global organisation with a multicultural workforce being able to answer questions in the chosen language of one of his team members could be very interesting. Watching Zuckerberg's clone development closely may tell us how much this technology can be trusted.

RunwayML has free credits to try this new tool out. To set up an avatar, navigate to characters through this link and use the ‘join a meeting’ button to see the AI come to life.  

Kudos to the developers at RunwayML who generously made my avatar look considerably younger.

🔥 In Case You Missed It… 44% Of Daily Music Uploads Are AI

A grey area exists over trying to game the system of the streaming platforms. The economic model is that the higher the share of streams, the higher the revenue. And these uploaded tracks are often aimed at ambient sound, sleep music, meditation music, etc. There is also some slightly more nefarious behaviour going on, uploading artist names similar to real artists and using keyword stuffing in an attempt to get picked up by the recommendation engines.

Streaming platforms want credibility of real artists or they fear that users will go elsewhere, but the dilemma is that we're seeing more success from AI-generated artists. IngaRose reached number one in multiple global markets. Breaking Rust hit number one on a Billboard chart in late 2025.
The digital industry has been plagued by bad actors creating click farms for ads, SEO content farms, and app store spam and this is one form of the AI iteration. Time will tell whether we are seeing consumers fully accept AI-generated music like IngaRose, Velvet Sundown and Breaking Rust. Until then the streaming platforms will continue to play whack-a-mole as bad actors find angles to exploit.

🏆 Tools, Podcasts, Products Or Toys We’re Playing With This Week

ChatGPT’s Image 2.0  

If your experience with AI image generation has been seeing people with extra limbs appearing in photos or hands with six fingers, take a look at the new offering from ChatGPT.
Tucked away in the release of GPT 5.5 is OpenAI’s new image model. DALL-E was OpenAI’s original app, a text-to-image model which over time has been integrated into ChatGPT's chat function.

I asked ChatGPT to include a drawing of an elephant I did when I was six to the Louvre Gallery in Abu Dhabi alongside famous artists. Not only is the result very impressive, but when you zoom in to the text that describes the pictures, it is sharp and accurate, which is something that GPTs have had a problem doing before.

A noticeable advancement in text-to-image models has been the editing function. Old models of AI image generation had this habit of throwing in new elements to an image when you tried to change one thing. For instance you ask for a table to be removed and two lamps appear on a nearby table at the same time.
GPT Image 2.0 is also great for product shots and infographics and in thinking mode it also has web search so you can pull real elements such as logos and product shots straight into the image. Movement is the only area where competitors are noticeably ahead.

To try out GPT Image 2, go to ChatGPT and click on 'create image', turning on Thinking Mode for anything complex.

Available both in the free versions (with limitations) and the paid versions. Hopefully this heralds the end of the examples of bad AI generated images that get published and should never have seen the light of day. Whilst new image generation may be good, we're still going to need humans in the loop to avoid such calamities in the future.

Did You Know? 

Till next time,

And The Rest Is Leadership