Large language models are having their Stable Diffusion moment

Large language models are having their Stable Diffusion moment. If you were around for the image-generation explosion of 2022-2023, you know exactly what I mean. If not, here's the short version: a technology that once felt like magic is becoming cheap, ubiquitous, and increasingly hard to differentiate. The same pattern is now playing out with text-based AI. And it's going to change how you think about content creation.

I've been building content workflows with AI tools since GPT-3 first became accessible. Back then, getting a coherent paragraph felt like pulling a rabbit out of a hat. Now? I can generate a decent blog outline in under 10 seconds using half a dozen different tools. The quality gap between the best model and the rest has shrunk dramatically. That's the Stable Diffusion moment in a nutshell.

But here's what nobody's talking about. When the models all perform similarly, the bottleneck shifts. It's no longer about which AI you use. It's about how you use it. And most people haven't figured that out yet.

What Actually Happened With Stable Diffusion?

Let me rewind for context. In 2022, DALL-E 2 was the undisputed king of AI image generation. OpenAI had built something genuinely impressive, and access was tightly controlled through a waitlist. Then Stability AI released Stable Diffusion as open-source. Anyone could download it. Anyone could run it locally. Within months, the market flooded with image generators—Midjourney, Leonardo, dozens of others—all built on similar underlying technology.

The result? Image generation became a commodity. Prices cratered. Differentiation moved from "who has the best model" to "who has the best user experience, workflow integration, and specialized features." Sound familiar? It should.

According to a 2024 analysis by Andreessen Horowitz, the marginal cost of generating AI content has been dropping by roughly 70-90% per year across both text and image modalities. That's not a typo. The economics are brutal for anyone trying to compete on model quality alone.

3 Signs LLMs Are Following the Same Trajectory

I've been watching this unfold in real time. Here are the three clearest signals that text-based AI is deep into its commoditization phase.

1. The open-source models are catching up fast. Meta's Llama 3, Mistral's models, and a wave of fine-tuned variants are now competitive with GPT-4 on many benchmarks. Not all benchmarks, sure. But enough that for most practical content tasks—blog posts, emails, product descriptions—you'd be hard-pressed to tell the difference in a blind test. I've run these tests myself. The results are humbling.

2. Price wars are everywhere. OpenAI, Anthropic, and Google have all slashed API prices multiple times in the past 18 months. DeepSeek's release in early 2025 sent another shockwave through pricing, with some providers offering access at literally 1/50th the cost of GPT-4 from a year ago. When the pricing curve looks like a cliff, you're in commodity territory.

3. The "wrapper" problem is real. Remember how every image tool was just Stable Diffusion with a nicer UI? The same thing is happening with text. Hundreds of startups have built thin wrappers around GPT-4 or Claude. They're competing on UI, templates, and marketing—not on core model capabilities. That's textbook commoditization.

Why "Which Model?" Is Becoming the Wrong Question

I get asked constantly: "Which LLM should I use for content?" My answer now is different than it was a year ago. A year ago, I'd say GPT-4, no question. Today? It depends entirely on your workflow.

Here's what I mean. If you're writing prompts manually in ChatGPT, switching to Claude might give you marginally better results on certain tasks. But the gain is small. Maybe 5-10% improvement on output quality, tops. What actually moves the needle is everything that happens around the model.

The prompt engineering. The content strategy. The editing process. The distribution plan. These things matter far more than whether you're using GPT-4o or Claude 3.5 Sonnet. I've seen terrible content produced by the best models and brilliant content produced by mid-tier models with great prompting. The model is not the differentiator anymore.

This is exactly what happened with cameras. For decades, people obsessed over megapixels and sensor sizes. Then smartphone cameras got good enough that the difference stopped mattering for 95% of use cases. The skill shifted to composition, lighting, and editing. Same pattern. Different technology.

The Real Bottleneck: Prompt Engineering Is Still a Mess

Here's where things get frustrating. The models are commoditizing, but the skill required to use them well hasn't kept pace. Writing effective prompts is still a specialized skill. Most people are terrible at it. I don't say that to be harsh—I say it because I've reviewed thousands of prompts from teams I've consulted with.

The most common mistakes I see:

• Vague instructions that produce generic output ("Write a blog post about marketing")
• Over-specifying to the point of strangling creativity (500-word prompts for 300-word outputs)
• No context about audience, tone, or purpose
• Treating the AI like a mind-reader instead of a tool

I've written about this extensively in my guide to writing AI prompts, but the short version is: most people spend more time choosing which AI to use than learning how to use it effectively. That's backwards.

And here's the kicker. Even if you do learn prompt engineering, it's still time-consuming. Crafting a detailed prompt for a complex blog post can take 10-15 minutes. Then you iterate. Then you edit. Suddenly the "time-saving" AI tool has eaten up an hour of your day. There's a reason so many people find ChatGPT prompts frustrating—the interface between human intent and model capability is still clunky as hell.

What the Stable Diffusion Moment Means for Your Content Budget

Let's talk money. When image generation commoditized, the smart move wasn't to buy the most expensive image tool. It was to invest in the workflow around it. Companies that built efficient pipelines for generating, selecting, and editing images got 10x more value than companies that just subscribed to Midjourney and called it a day.

The same logic applies to text. If you're paying $20/month for ChatGPT Plus and another $20 for Claude Pro, you're spending $480/year on raw model access. That's fine. But are you getting $480 worth of productivity out of it? For most people, the answer is no—not because the models are bad, but because the prompting bottleneck eats up all the efficiency gains.

I've found that the teams getting the most ROI from AI content are the ones who've either (a) invested heavily in prompt engineering training, or (b) moved to tools that handle the prompting for them. Option A works but takes time. Option B is where the market is heading.

This is the natural evolution of any commoditizing technology. The value moves up the stack. Raw model access becomes cheap; the ability to extract useful output from those models becomes the premium offering. It's why zero-prompt AI content generators are gaining traction—they solve the actual bottleneck, not the one that existed two years ago.

4 Things You Should Stop Worrying About

Given where we are in the commoditization curve, here's what I'd stop obsessing over if I were you.

1. Benchmark scores. Unless you're doing academic research, the difference between a model that scores 86.5% and one that scores 88.2% on some benchmark is irrelevant to your content workflow. Real-world performance rarely tracks benchmark performance perfectly.

2. Which company has the "best" model this month. The leaderboard changes constantly. By the time you've switched your workflow to the new leader, someone else has probably released something competitive. Pick a tool that works and focus on using it well.

3. Whether AI content is "good enough" yet. This debate is tired. AI content is good enough for many use cases and not good enough for others. The question isn't whether AI can write like Hemingway. The question is whether it can handle your specific content needs. Test it. You'll know in 20 minutes.

4. Model loyalty. I see people who refuse to try anything except ChatGPT, or who switched to Claude and won't look back. This is like being loyal to a hammer brand. Use the tool that works for the job. The models are converging anyway.

What You Should Actually Focus On Now

So if the model doesn't matter as much anymore, what does? Three things.

First, your content strategy. AI can write words. It can't decide what words are worth writing. That's still your job. The companies winning with AI content are the ones with clear audience understanding, strong topic selection, and distribution plans that go beyond "publish and pray."

Second, your editing process. AI output is a starting point, not a finished product. The difference between mediocre AI content and great AI content is almost always the human editing step. I spend roughly as much time editing AI drafts as I used to spend writing from scratch. The difference is that the final product is better, not that I'm working less.

Third, your workflow efficiency. This is where the Stable Diffusion parallel really bites. When the model is a commodity, the interface becomes everything. How fast can you go from idea to published content? How much friction is in your process? Every extra step between "I need a blog post" and "it's live on my site" is costing you money.

This is why I've been experimenting with tools that collapse the workflow. AI-Mind, for instance, lets you skip the prompt-writing step entirely—you describe what you need, pick a content type, and it handles the rest. The first 30 generations are free, which makes it easy to test whether the zero-prompt approach actually saves time. In my experience, it cuts the "staring at a blank prompt box" phase down to about 30 seconds. That adds up.

There are other approaches too. Some teams build elaborate prompt libraries. Others train custom models on their brand voice. The right answer depends on your volume and budget. But the common thread is this: stop worrying about the model. Start worrying about the system around it.

Key Takeaways

• LLMs are following the same commoditization path as image generators—model quality is converging and prices are in freefall.
• The question "which model is best?" matters far less than "how efficiently can I get useful output from any model?"
• Prompt engineering remains the biggest bottleneck in AI content workflows, and most teams underinvest in solving it.
• Value is shifting from raw model access to workflow tools that reduce friction between human intent and AI output.
• The teams winning with AI content focus on strategy, editing, and workflow efficiency—not on chasing the latest model release.

Sources

Andreessen Horowitz, The Economic Case for Generative AI, 2024. Analysis of declining inference costs and commoditization trends across AI modalities.

Stability AI, Stable Diffusion Public Release Announcement, 2022. The open-source release that triggered the image generation commoditization wave.

Artificial Analysis, LLM Quality and Price Comparison, 2025. Independent benchmarking tracking the convergence of model quality and the decline in API pricing across major providers.