OpenAI Calls a ‘Code Red’ + Which Model Should I Use? + The Hard Fork Review of Slop

That's really important, and I want to underscore that. What's happening here is a combination of factors. For a while, OpenAI and, to a lesser extent, Anthropic, were surviving on the moat of their models. They had the best models globally, which separated them from the rest. If you wanted to work with a world-class model, you were likely doing agentic software development or vibe coding. You needed the smartest model possible and were willing to pay a premium for access. Your alternative was Llama, Gemini, or other second-rate models.

If OpenAI fails, we will easily identify 15 major mistakes they made. However, it's equally possible that their current bets will pay off. We are currently in a period of uncertainty. For those presenting the bear case this week, consider this: the company is massively leveraged. They've committed trillions of dollars in spending, relying on revenue that is far from materializing. Furthermore, their product organization lacks focus, attempting a little bit of everything.

I have a really good one. My boyfriend works at Anthropic, so take what I say with an 80% discount. Before 4.5, I wasn't using Claude daily, though I tried it occasionally like other models. ChatGPT and Gemini were my daily drivers. When Opus 4.5 launched, I tested it with an unpublished study, asking it to write a column in the style of Casey's Platformer. ChatGPT 5.1 still fails this, giving bullet points I'd never use. Gemini 3's output is somewhat structured like mine but has obvious AI tells. When I first tried this with Opus 4.5, it sent a chill down my spine. For the first time, I saw sentences I could have written, especially the conclusion. Earlier this year, we discussed style transfer, like the Studio Ghibli moment for images. I've been waiting for that in text, and this was it. I thought, "Oh my God, it's starting to happen." That was the first thing Opus 4.5 did that made me think they might have something here.

Currently, Anthropic's primary focus is the enterprise. At the start of this year, Anthropic had less than a billion dollars in annualized revenue. By year-end, they expect to reach approximately $9 billion in annualized revenue. This growth is largely due to selling into the enterprise market. If you're a developer or a large consulting firm looking to create agentic workflows, many companies are purchasing this software from Anthropic. Consequently, Anthropic has become one of the fastest-growing startups ever, creating a massive opportunity. If Anthropic weren't a player, that $9 billion would likely go to competitors like OpenAI and Google. This represents a significant amount of revenue OpenAI is missing out on this year. OpenAI projected $20 billion in revenue, so imagine how different their situation would be if they had captured the enterprise market. Increasingly, Anthropic is dominating this space.

The argument made was a critique of ChatGPT, stating it "sucks" because it's merely an amalgamation of internet content, lacking soul. I considered the metaphor of a blurry JPEG. When I used Opus this week and Gemini 3 the week before, I had the sensation of a webpage loading a JPEG: initially blurry, then gradually resolving to higher resolution. We are now in a moment where AI is achieving higher resolution. That was my feeling when Claude created something with sentences that, for the first time, felt like me. It was as if the blurry JPEG was becoming less blurry. Therefore, I cannot give a single answer to which model you should use, as the answer will consistently change over the next six months to a year. If you truly care about this, you'll need to try new things.

I've decided on a principle for my life: I will not listen to opinions about AI from people who do not use it. If you lack firsthand experience with these models, having spent at least five to ten hours with the newest versions, you are discussing something that no longer exists.

I believe both scenarios can be true. We are on a trajectory where AI will likely solve coding and software engineering first. While software engineers will still exist, they won't be writing code manually. On our predictions episode later this month, I might predict that by the end of 2026, coding will be effectively solved, with many tools, even free ones, handling it. However, many other jobs remain. Significant translation work is still needed, and not every job has as defined a structure as coding. So, it's true that models are advancing, bringing us closer to automating software engineering. Yet, for professions like accounting, law, or medicine, AI is currently only momentarily useful. The question then becomes: what will it take to generalize the solution for coding to every other job, and how long will that take?