Skip to content
followmy.ai
Blog

OpenAI's New Models Are Here, and People Are Mad. Should You Bother?

OpenAI's new GPT-6.2 Astra and Sol models are facing backlash from power users, raising the question of whether chasing the latest AI update is worth the troubl

By Craig Mason 8 min read

The short version

OpenAI just dropped two new models, GPT-6.2 Astra and Sol. The company says they are better and, in one case, cheaper, but many power users I follow are calling them a major step backward for creative and complex work. This whole update cycle makes me seriously wonder if I should just ignore the hype and stick with the tools I know are reliable.

So, what did OpenAI actually announce?

On the surface, it looks like a standard refresh of their top-tier models. OpenAI released two new flagship models: GPT-6.2 Astra and GPT-6.2 Sol. The company has positioned Astra as the powerhouse for highly complex professional tasks, the kind of reasoning and analysis you might use for scientific research or deep financial modeling. GPT-6.2 Sol, on the other hand, is framed as the efficient workhorse for more common, everyday professional work. Think drafting emails, summarizing meetings, and generating code snippets.

You can access these new models in a few places if you have the right subscriptions. They are available in ChatGPT Work, the specialized version for businesses, as well as in Codex for developers. Of course, they are also accessible through the API, which is where the most serious developers and businesses build their applications.

OpenAI also adjusted the pricing, which is always a key part of these announcements. For GPT-6.2 Astra, they’ve lowered the price by 20% compared to the model it replaces. This move puts its cost directly on par with its main competitor, Anthropic’s Claude Fable 5.2. Price wars are good for us customers. GPT-6.2 Sol’s pricing, however, remains the same as its predecessor. The announcement was paired with the usual charts showing improved scores on industry benchmarks, like the Agents’ Last Exam, and claims of better token efficiency. More power for less money. Sounds great.

Why are people saying these new models are ‘unusable’?

Because for a very vocal and influential group of AI power users, the new models are failing at the exact tasks they were built for. Almost immediately after the launch, the backlash started. The initial report on the launch captured this frustration perfectly, citing feedback that was anything but positive. Developers and AI creators took to social media to share their experiences, and the reviews were brutal.

Words like “essentially unusable” and “the worst model OpenAI ever released” began circulating. These were not coming from casual users asking for recipes. These critiques came from people who build AI-powered agents and rely on the models for high-stakes creative and delegated tasks. Matt Shumer, who builds autonomous agents, was one of the most prominent critics. He found that the new models struggled to follow complex instructions and failed at tasks that older models like GPT-4 handled with ease. His experience pointed to a significant regression in the model’s ability to act as a reliable, autonomous worker.

The complaints seem to cluster around a specific type of failure. The models are reportedly worse at maintaining context over long conversations, less creative in their outputs, and more prone to misunderstanding nuanced instructions. For anyone trying to build a product that delegates work to an AI, this kind of unpredictability is a disaster. It means more time spent re-prompting, correcting, and cleaning up the AI’s mistakes, which defeats the entire purpose of using it. This is a very different problem from a model simply being wrong about a fact; it’s about the model failing to be a useful tool.

Is GPT actually getting dumber?

This is the question that seems to pop up every single time a new model is released. It’s become a recurring part of the AI conversation. A new GPT version arrives, and a segment of the user base insists it has been secretly “lobotomized.” There might be several things happening at once here, and none of them are simple conspiracies.

One strong possibility is that the process of making models safer also makes them less capable. As OpenAI fine-tunes its models to refuse harmful or inappropriate requests, that alignment process can sometimes overcorrect. The model becomes so cautious and agreeable that it loses its creative spark and its ability to handle edgy or ambiguous prompts. It feels like its personality has been sanded away, leaving a sterile and unhelpful assistant in its place. The model becomes less likely to say something wrong, but also less likely to say something brilliant.

Another factor could be the optimization techniques used to make these massive models faster and cheaper to run. Methods like quantization, which reduces the precision of the numbers used inside the model, can save huge amounts of computational power. While these techniques are designed to have minimal impact on performance, they aren’t perfect. For the vast majority of users, the difference is unnoticeable. But for those pushing the models to their absolute limits, these tiny degradations can add up to a noticeable drop in quality for the most demanding tasks. The model may ace a multiple-choice benchmark but fail a qualitative test of reasoning.

Finally, our own expectations play a huge role. As we get more accustomed to working with these tools, we get better at spotting their flaws. What seemed magical a year ago now feels clunky. We are constantly pushing the boundaries of what we ask the AI to do, and our standards are rising faster than the technology is improving. It’s possible the models are not getting dumber, but that we are simply getting smarter about how we use them, which makes their existing limitations all the more obvious.

How does this launch compare to past OpenAI updates?

This cycle feels incredibly familiar. The launch of GPT-6.2 Astra and Sol is not an isolated incident; it follows a pattern we’ve seen with previous OpenAI updates. When GPT-4 Turbo was first released as a faster, cheaper successor to the original GPT-4, the community response was almost identical. A wave of initial excitement was quickly followed by a chorus of complaints from developers and advanced users who felt the new model was a significant downgrade in reasoning and instruction-following.

I remember the posts and the arguments well. People shared side-by-side comparisons showing how GPT-4 Turbo would miss the nuance in a prompt that the original GPT-4 understood perfectly. The consensus among many was that in the race to make the model faster and more cost-effective, OpenAI had compromised on the raw intelligence that made GPT-4 so impressive in the first place. Sound familiar?

Over time, OpenAI seemed to address some of those concerns with subsequent updates to the Turbo model. The company never explicitly said, “You were right, we made it dumber and then fixed it.” Instead, they released new preview versions and eventually a final production model that seemed to have regained some of its lost capabilities. This history suggests that the initial version of a new model might not be the final word on its performance. There is often a period of public testing and feedback that leads to further refinement. The models we are using today might be a rough first draft, not the finished product.

What should I actually do with Astra and Sol?

My advice is simple: proceed with caution. Do not rush to tear out the engine of your existing workflows and replace it with Astra or Sol, especially if you have a system that is working reliably for you right now. The potential for a 20% cost saving with Astra is tempting, but that saving is worthless if you have to spend hours debugging prompts or your application’s quality drops.

Instead, treat these new models as experimental options. Set up a parallel workflow. Create a testing environment where you can put Astra and Sol through their paces using your own real-world tasks. The most important thing is to test for your use case. Don’t rely on benchmarks or social media chatter alone. Your prompts are unique. Your needs are specific.

A great way to do this is to take a prompt that you know works perfectly with an older model like GPT-4 Turbo. Run that exact same prompt through Astra and Sol ten or twenty times. Log every single output. Then, sit down and compare them. Is the quality consistently as good? Is it better? Is it more erratic? Are you seeing the kinds of failures that others have reported? This kind of head-to-head testing is the only way to know for sure if the new models are right for you. It takes time, but it will save you from a production crisis down the road.

Is it better to just stick with what works?

This is the core question for anyone who uses these tools for their job, and honestly, my answer is increasingly yes. The constant pressure to be on the cutting edge is exhausting. For a while, every new model was a huge leap forward, and adopting it was a no-brainer. That is not the case anymore. We seem to be entering a phase of smaller, incremental gains, and sometimes, even regressions.

For my own work, particularly projects for clients, reliability is everything. I would much rather use a slightly older, more expensive model that I know I can count on than a brand-new model that might be faster on paper but is unpredictable in practice. The cost of a failed task or a garbled output is always higher than the few cents I might save on an API call. An AI that you have to constantly supervise is not much of an assistant.

There is a real business cost to this endless churn. Every time a new model drops, it requires time for testing, re-evaluating prompts, and potentially re-writing code. That is time not spent on building products or serving customers. So, my personal strategy is shifting. I’m no longer jumping on every new model the day it’s released. I’m going to wait a few weeks, let the community establish a clearer consensus, and do my own careful testing before I even consider making a switch for anything important. Sometimes, the smartest move is to let others ride the bumpy frontier and just stick with what works.

FAQ

What are GPT-6.2 Astra and Sol? GPT-6.2 Astra is OpenAI’s new model for complex, professional reasoning tasks, while GPT-6.2 Sol is designed for more common, everyday work. They are the latest flagship models from the company.

Are the new OpenAI models really worse? It’s complicated. OpenAI’s benchmarks show they are more capable, but many early power users report that they are less creative and perform worse on complex, agent-like tasks compared to older models.

How much do GPT-6.2 Astra and Sol cost? GPT-6.2 Astra is priced 20% lower than its predecessor, making it competitive with Anthropic’s models. The pricing for GPT-6.2 Sol remains the same as the previous version.

Should I update my apps to use the new models? It’s best to test them thoroughly first. Do not switch any mission-critical systems to the new models until you have verified that their performance is acceptable for your specific use cases.

Found this useful? Read more from the blog →