Two years ago, the choice between an open AI model and a paid one wasn't really a choice. The paid models were simply far better.
That's no longer true, according to someone with a front-row seat. On IBM's latest earnings call, CEO Arvind Krishna told investors that open-weight models have become good and cheap enough to make serious sense for businesses running AI on their own hardware.
His numbers tell the story. In 2024, on a broad benchmark, a top closed model might have scored around 70 while the best open model scored 20 or 30. By late 2025, the gap had narrowed to roughly 70 versus 50. Today, he says, open models are about 90% as effective as frontier models, at somewhere between a tenth and a quarter of the token cost.
His framing for businesses was simple: would you pay ten times more for the last 10% of quality?
The paid models are already responding
If open models were creeping up quietly, the paid AI companies would have no reason to react. They are reacting.
At the end of July, OpenAI cut the price of its GPT-5.6 Luna model by 80% and Terra by 20%. In August, it cut its frontier GPT-5.6 Sol model by more than 20% for three months, amid growing competition from Anthropic and Chinese AI labs. Then came GPT-6 Sol and Luna, priced at half or less of what the previous versions cost, a cut OpenAI says is permanent.
Anthropic took a slightly different route. In July, it released Claude Opus 5, keeping the same price as its predecessor but claiming performance close to its most expensive model at half that model's price. Two months later, Opus 5.5 arrived with a further 20% price cut.
Forbes described OpenAI's cuts as a sign that competition from open-weight models is forcing proprietary providers to lower prices. Whatever the exact cause, the direction is clear. Prices are falling on both sides, and fast.
Why "good enough" matters so much
The key to Krishna's argument isn't that open models are the best. They aren't. It's that most work doesn't need the best.
He pointed to everyday tasks like summarizing an email, reviewing a contract, answering a customer call, or drafting a message. Some tasks genuinely need the top models, he said, but the vast majority can be handled just as well by the cheaper option.
This connects to a bigger pressure. Apollo's chief economist recently showed that AI hasn't yet paid off for most companies outside the tech sector, and that many are starting to treat token spending as a cost that hurts margins. Forrester has predicted that companies will lean more on cheaper open models, saving expensive ones for critical work.
When businesses are under pressure to justify every dollar, "90% as good for 10% of the price" is a very hard offer to ignore.
The case for open models taking over
Supporters of open models have several strong arguments.
The price gap is huge. Even after all the cuts, paid frontier models still cost several times more per token than running a capable open model at scale. For companies processing millions of requests, that difference adds up quickly.
Control and privacy. Running a model on your own hardware means your data never leaves your building. For banks, hospitals, governments, and anyone handling sensitive information, that can matter more than a few points on a benchmark.
No surprise price changes. Paid model pricing is now a moving target. Some of OpenAI's cuts are permanent, while others are promotions with an end date. A company running its own open model knows exactly what its costs will be next year.
History has examples. In servers, Linux eventually won much of the market from paid operating systems, not because it was always better, but because it became good enough, free to adopt, and impossible to lock anyone into.
The case for paid models staying ahead
The other side has strong arguments too.
The frontier keeps moving. "90% as good" is measured against today's best model. Paid labs release new models every few months, so the target open models are chasing keeps moving. Being 90% as good as last season's leader may not be enough for the hardest tasks.
The gap is closing from both directions. Open models are getting better, but paid models are getting cheaper. If paid models keep cutting prices, the cost advantage of open models shrinks, and the convenience of paid models becomes more attractive.
Free to download isn't free to run. Krishna himself admitted the catch. Many companies will struggle to run these models, he said, because they lack the specialist engineers needed for GPUs, networking, and performance tuning. Hardware, electricity, and skilled people all cost money. For many businesses, a paid API with no infrastructure to manage may still be cheaper overall.
Convenience usually wins. Cloud computing grew even though running your own servers was often cheaper on paper. Many companies were willing to pay more to avoid the hassle.
Everyone has something to sell
It's worth noticing who is making each argument. IBM sells exactly the kind of on-premise infrastructure that open models need, and Krishna expects on-prem AI demand to boost sales of IBM's own hardware. OpenAI and Anthropic sell access to paid models through their APIs. Every voice in this debate has a business interest in the outcome.
That doesn't make anyone wrong. But it's a good reason to look at the numbers rather than the pitch.
Where this could go
There are several possible futures here.
Open models could become the default for everyday work, the way Linux did for servers, leaving paid models as premium tools for the hardest problems. Paid models could keep cutting prices until the savings from open models no longer justify the effort of running them. Or companies could end up mixing both, sending simple tasks to cheap open models and harder ones to paid frontier models.
Right now, nobody knows which one will win. What is clear is that the price war has started, and the companies paying the bills are the ones who will decide how it ends.
So I'll ask you: if you're building with AI today, are you using open models, paid models, or both, and what made you choose?
Sources
- The Register, IBM CEO says open weight models are now irresistibly good and cheap on-prem, October 22, 2026
- Technology.org, OpenAI Cuts GPT-5.6 Sol Developer Pricing by More Than 20%, August 2026
- Yahoo Finance, OpenAI cuts GPT-6 Sol and Luna API prices by 50%, 2026
- Let's Data Science, Anthropic Releases Claude Opus 5 at Lower Cost, July 2026
- BetaNews, Anthropic launches Claude Opus 5.5, September 2026
- Forbes, Why OpenAI's 80% Price Cut Could Trigger A Race To The Bottom In AI, July 31, 2026
- Apollo, No Signs of AI Payoff Outside the Tech Sector, October 24, 2026
- Computerworld, AI devours a fifth of companies' IT budgets, October 2026
Top comments (10)
The real advantage isn’t just having access to cheaper AI models. It’s knowing which tasks don’t need the most powerful model in the first place.
As the gap in cost keeps growing, the teams that understand where to spend and where to save could have a serious edge.
Agreed, knowing where good enough is good enough may become a skill in its own right.
One twist: with paid models cutting prices too, the cost gap might shrink rather than grow.
Either way, the teams that test on their own tasks will spot the shift first.
I think the future is hybrid. I use both paid models and open models depending on the task, but I believe the real advantage is not choosing the most powerful model every time; it is knowing when a cheaper model is good enough and when a more capable model is worth the extra cost.
For developers building real products, the total cost of ownership matters just as much as model performance. A cheaper model that requires significant infrastructure and maintenance may not always be cheaper in practice.
The exciting part is that this competition gives developers more options, reduces vendor lock-in, and makes AI more accessible to smaller teams and independent builders. Ultimately, the winner won't necessarily be the model with the highest benchmark score, but the one that delivers the most value for a specific real-world problem.
Hybrid is probably where many teams land first. The hard part is the deciding itself: knowing which tasks are good enough on a cheap model takes testing on your own work, not benchmark tables. Teams that build that habit will save the most. And you’re right, total cost of ownership is where cheap and expensive often trade places, both ways.
I completely agree. Testing models against real-world tasks rather than relying solely on benchmarks is such an important point, and I think there's a lot to learn from your perspective on this.
You clearly have valuable knowledge and experience in this space, and I'd love to learn from your insights as I continue growing as an AI engineer. I'd be glad to connect and keep learning from your work and discussions here on DEV!
Thank you, that’s very kind. I’m still learning too, which is half the reason I write these posts. Happy to connect, and I’d love to hear what you discover in your own testing. Good luck with the AI engineering journey!
In our projects, we use a hybrid approach.
For repetitive, internal tasks with high data volumes (such as entity extraction, support ticket summarization, or text classification), we run open-source/open-weight models (Llama or Qwen) hosted on our own infrastructure. It turns out infinitely cheaper at scale, and we don't have to worry about GDPR or confidential customer data.
That’s a great real-world example of the hybrid split, and those tasks fit open models perfectly.
High volume, clear outputs, and sensitive data is exactly where self-hosting pays off most.
I’d be curious how the infrastructure and maintenance costs compare once you count engineering time.
And what still goes to paid models, the tasks where the cheaper option fell short?
The "would you pay ten times more for the last 10%" question assumes the missing 10% is spread evenly. I'd guess it bunches up in a few task types, so I'd run the open model on my own task mix and find where it fails before routing anything to it. Email summaries are an easy call, but which side would you put contract review on?
Pls check my articles♥️