[AI SPRINT] OpenAI's GPT-4.5 and 6 AI Vendor Red Flags

GPT-4.5 is Here: But Bigger Isn’t Always Better

The latest OpenAI release, GPT-4.5, has arrived—and it's making waves, but not entirely for the right reasons. Despite being the most expensive model yet (2900% pricier than GPT-4o and 2000% above Grok3), its benchmark results and real-world performance have been surprisingly underwhelming. Many now question whether ultra-massive models represent AI’s future or if the path forward lies with more focused, cost-effective alternatives pioneered by DeepSeek and others.

What's particularly puzzling is OpenAI’s marketing emphasis on GPT-4.5’s "vibe" and emotional intelligence rather than groundbreaking technical abilities. OpenAI argues that extensive training has provided the model with unmatched understanding of human intent, allowing more fluid and intuitive interactions. That sounds great, but personally, I don’t think this is the biggest challenge with AI models today.

So this leaves me wondering: have we reached a point of diminishing returns with large language models (LLMs)? Are incremental improvements now disproportionately expensive without delivering transformative leaps?

My Take: LLMs May Be Good Enough for What They Need to Do

While GPT-4.5 might not seem revolutionary at first glance, it will likely replace GPT-4o as OpenAI’s standard general, non-reasoning model. Its extensive knowledge base, enhanced contextual awareness, and improved interaction tracking mean it's well-suited as an orchestration layer in AI workflows rather than just another standalone chatbot.

But the elephant in the room remains cost. The massive compute requirements associated with GPT-4.5 raise critical questions: is scaling models infinitely upward the best strategy anymore? Competitors like DeepSeek, Perplexity, and Anthropic’s Claude are increasingly demonstrating that targeted, efficient innovations can deliver more substantial and affordable improvements.

More significantly, we may be reaching the limits of this generation of general-purpose language models. Current models offer impressive conversational flow, broad knowledge bases, and solid contextual comprehension. The real opportunities for advancement likely lie elsewhere: in improving reasoning capabilities, enabling multi-agent collaboration, reducing hallucinations, enhancing multimodal interactions, and achieving real-world task execution—not simply growing larger models.

So What Should You Do About GPT-4.5?

Here is what you need to know about where GPT-4.5 fits into the AI equation:

• Availability: Currently limited to only API and ChatGPT Pro users, but GPT-4.5 will soon roll out to Plus, Teams, Enterprise, and Educational users.

• Performance Metrics: While it surpasses GPT-4o in science, math, and hallucination avoidance (with real-world outcomes varying), it notably underperforms dedicated reasoning models, and often isn’t great at coding either.

• Better Conversationalist: Testers preferred GPT-4.5 roughly 60% of the time over older models, depending on task. This seems to be the top selling point.

• Multimodal Capabilities: Voice and multimodal features aren't yet included

But the real question is: do you need it? GPT-4.5 is interesting, but is a “Research Preview” right now. It won’t be revolutionary for most people, so unless you are an enthusiast, or your workflows heavily depend on nuanced emotional intelligence, detailed context tracking, and human-like interaction, it’s probably best to use other models for the moment, and in particular, Claude 3.7. Once it becomes fully available I expect it to become the primary AI model for ChatGPT, but that will be gradual over the next 6 months.

And GPT-5? I expect when it does come, it will be a hybrid AI coordinating across other AIs: using GPT-4.5 as the general purpose model, o3 for reasoning and deep research, Dall-E for images, etc.

Comparing Claude 3.7: The Cost-Effective Challenger

Anthropic’s Claude 3.7, launched shortly before GPT-4.5, has received significantly more positive feedback. Praised for coherence, improved coding, rapid responses, and stronger reasoning, Claude 3.7 provides superior product at a much lower price. It includes their new “hybrid thinking” model, bringing self-reflection in to ensure the AI completes the intended task properly before it gives you results. Optimized for coding tasks, it solidifies Claude as the best copilot for software development, among many other tasks. Unfortunately, it still lacks web searching and some other capabilities. For businesses, this means that it’s best to still standardize on ChatGPT if you want to have the most bang-for-the-buck on capabilities and performance, and then supplement with Claude or Perplexity as needed.

What’s Next for AI?

The lukewarm reception of GPT-4.5 strongly suggests we’re approaching the practical limits of what current transformer models can achieve through sheer scale. Now, AI is at a crossroads: should we keep pushing model size, or shift toward innovative architectures, specialized models, and multi-agent systems?

One thing is certain—the next breakthrough won’t come simply from more compute. The highest returns will come from improving how models reason, collaborate, and execute real-world tasks, marking the true frontier of AI innovation.

Navigating the AI Gold Rush

Now, on to how to choose an AI vendor to support your AI transformation. As you know, AI is the hottest topic in business today, and vendors are scrambling to cash in on the gold rush. Over the past few months, I’ve met with more than 25 AI software and service providers, and what I uncovered should be a wake-up call for any business leader considering AI adoption.

The reality? Many providers make big promises but few have actually delivered generative AI applications.

On average, the firms I met with had completed only 5 generative AI projects in production with clients in the last 12 months, with a variety of prototypes that never panned out. This spans from newly established firms (what I call AI Prototype Labs) to even large, existing software development firms adding generative AI as an expertise.

The summary: Everyone is learning these new tools, so investing now is still a little risky, and you are best served to seek out companies that have real, demonstrated experience with AI software development and machine learning capabilities.

To help navigate this, I’ve written a 30-page report breaking down provider types, capabilities, red flags, and recommendations on who to trust with your AI investments.

Today I’m sharing a summary of what you need to watch out for—and how to find a vendor that truly meets your needs. 

The 6 Red Flags in AI Services Today

Across everyone I’ve met with and seen in the market so far, here are the biggest red flags leaders should watch out for when hiring an AI vendor:

1. Everyone Is an "AI Expert"

You’ve seen it all over LinkedIn: suddenly, everyone is an AI expert. But dig deeper, and many have little hands-on experience building real AI systems. With so few actual projects being completed, few have experience building generative AI-based applications. Further, many “AI strategists” lack real business strategy or innovation expertise, instead just knowing how to use off-the-shelf generative AI applications. Even outsourced IT firms and software developers branding themselves as "AI firms" often have minimal true “AI” expertise.

Pro Tip: Always ask for qualifications and references to verify actual AI technical experience, and look for experience with machine learning in addition to generative AI.

2. Few Deployed GenAI Projects

The average number of generative AI projects fully deployed by providers in the past year? Five. This was one of the stunning revelations: even well-established, larger firms have little actual experience delivering production-ready generative AI solutions. Most are still experimenting, building prototypes but not maintaining or scaling AI solutions. The most knowledgeable AI providers are actually not the run-of-the-mill AI software development shop, but AI platform startups who need to help with onboarding customers, like https://credal.ai. AI Platform providers have the best insights and visibility into current AI adoption, are invested into ensuring long-term support for their clients, and are definitely worth a detailed look when building any AI application.

Pro Tip: Always ask for references and case studies to confirm a provider’s real-world experience.

3. Unclear IP Ownership

Many AI vendors retain ownership over the solutions they build—often without making it clear to clients. Whether due to their investment in research, reusable AI components, or plans to resell solutions, this can leave businesses without full control over their own AI. Worse, you may be subsidizing development of a product they intend to sell to your competitors.

Pro Tip: Read contracts carefully and ensure you retain full IP ownership over the AI developed for your business.

4. Hidden Long-Term Costs

AI is never a one-and-done project. AI models degrade over time and require continuous updates to remain accurate as new models and capabilities emerge. Plus, the more you use AI, the more operational costs increase. Many vendors downplay this, leading to surprise expenses post-launch.

Pro Tip: Get a clear estimate of ongoing maintenance, support, and operational costs before committing.

5. Hidden Vendor Lock-In

Cloud and AI vendors, especially Microsoft, are paying huge financial incentives to AI service firms that push their platforms. Some AI vendors use platforms based on what makes them the most money—not what’s best for their clients. Worse, most don’t pass these incentives on to customers.

Pro Tip: Ask which platforms vendors use and what financial incentives they receive for recommending them. 

6. Freelancers and Startup AI Labs Lack Stability

While low-cost freelancers and small AI startups labs may seem attractive for small projects, they often lack security standards, development best practices, and financial stability. My guess is that many won’t be around in a year or two, leaving you with an unsupported AI system.

Pro Tip: Work with 20-plus person vendors that have a proven methodology and track record of long-term support and compliance.

How to Find a Reliable AI Vendor

If you’re serious about building AI into your business today, follow this structured approach to selecting the right provider:

✅ Start with Strategy First – Before jumping into AI software, engage a qualified AI Strategy and Transformation firm to define your business needs, assess data readiness, and create a roadmap. This is more than just figuring out ways to use AI: you need to start with identifying the top use cases, ideally focused on highest-return and lowest-risk technologies. Strategists may have partner firms they can connect you with (as I do).

✅ Vet Providers Carefully – Look beyond marketing hype. Verify AI expertise, check for unbiased platform recommendations, and confirm data security and compliance standards. Ensure they have machine learning expertise, and have deployed generative AI applications in a production capacity—not just a prototype.

✅ Clarify Ownership and Costs – Ensure you own all IP for your AI system, understand long-term costs, and avoid proprietary platform lock-in.

✅ Plan for Ongoing AI Maintenance – AI models degrade over time and require constant updates as AI capabilities improve. Make sure your vendor provides cost estimates for continuous monitoring, updates, and performance optimizations.

The AI services market is flooded with providers looking to make a quick profit. By recognizing the red flags and asking the right questions, you can avoid costly mistakes and find a vendor that truly supports your AI journey.

Want to dive deeper into AI strategy and vendor selection? Reach out—I’ve got the insider knowledge to help you navigate the AI landscape. I’m happy to share the full whitepaper.

Did someone forward this newsletter to you? If you're not already signed up, you can subscribe to AI SPRINT™ for free here.