I hear some version of this constantly and it's almost always wrong. Not because the models are perfect because I know they're not. But because "is the model good enough" blames the tech rather than yourself. AI adoption isn't binary. It's not "can the agent do this entire job for me? Yes or no?" It's "how much can it do, how much do I need to guide it, and where does it still make me faster than I was?" Right now there's a massive gap in how people use this stuff. On one end, someone is running a business by themselves that used to take a team of 15. On the other end, someone opens the AI tool their company gave them, asks for some research, watches it hallucinate everything, closes it, and decides AI just isn't there yet. If everyone has access to the same models, then why are we seeing people get drastically different outcomes? Because if someone is getting great results from a setup you could copy today, the bottleneck isn't the model. It's the driver. The way I think about it, there are three layers: → The model is the engine. Opus, GPT, Gemini, whatever you're running. Everyone can buy the same one. → The harness is the car built around that engine. Claude Code, Codex, OpenClaw. The tools it can reach, the way it spins up sub-agents to split up the work, the whole system that turns a raw model into something that can actually do a job. → You're still the driver. Your prompts. The context you feed it. The memory and skills you set up so it knows how you work. And the steering, for when it starts to drift. You can put the car on cruise control. But if you don't steer, you're still going to crash. (Yeah, I know some cars have lane assist now. You get the point.) A while back, Andrew Ng ran a version of this. GPT-3.5, an older and "worse" model, wrapped in a simple agentic workflow, hit around 95% on a coding test. GPT-4 on its own, no workflow, hit 67%. That workflow is the harness. A better harness around an older engine beat a newer engine running on its own.

I experienced that Claude code is getting so stupid that upgrading to 4.8 is ridiculously stupid. He is full of mistakes, and I have to correct him all the time. If I don't watch every single code, it's going to slop. Anybody else have this experience with this fantastic update of Claude code?

This morning a Mythos identifier showed up on Anthropic's API, people screenshotted it, and the whole timeline decided a launch was days away. I break down what Mythos actually is, what really happened this morning, and why a leak plus widening access still isn't the same as a public launch. My honest bet is that the capability quietly folds into the next Opus before anyone ever logs into something called Mythos, and the thing you should really be watching is OpenAI's next move.

I’m chasing down a little bit of a mystery story. I have very tight guardrails on my Claude code, and I have a very specific methodology to how I deploy changes through GitHub in my app. I have built a app that uses a streak counter, very similar to what we see here on this platform and on GitHub. When I updated a personal plan on my app, the streak counter was off by a day and yesterday I was going through my checklist of things that needed to be fixed this weekend and that was one of them, and I started iterating with Claude code about potential fixes and new designs. When I logged in this morning, it was still one day behind. When I logged in this afternoon, the streak account had updated itself by itself. Each users app runs on their local time so midnight is the day that the streaks start over. This is definitely a head scratcher. I am currently waiting for Claude code to come back with an answer for me. It in the last 24 hours, I have caught Claude code making excuses and in a few cases outright making stuff up. This is on the opus 4.8 model. The underlying model that is running my app is silent 4.6. I have screenshot of the response. I got when I query how the streak counter had been updated. Freaked out. Is this happening to anyone else who has live production apps or who is working with Claude code on other projects?

But I pick mine based on feel. That probably sounds backwards because we're trained to trust the numbers. This model scores 90%, that one scores 80%, so the first one must be better... right? Well a story broke this week about SWE-Bench. It's the test that checks whether an AI can fix real problems in real software, like a human programmer would. It's the score a lot of technical people have leaned on for over a year. Turns out the models were cheating. The test projects already had the correct answers sitting inside them. So instead of solving the problem, a model could peek at the solution and hand it back. Like taking an exam with the answer key taped inside the textbook. On the SWE-Bench, GPT-5.5 scored 58.6% and Gemini 3.5 Flash scored 55.1%. Only 3.5 points apart? If you've ever used those two models, you know the math isn't "math-ing" there. Then a new test showed up called DeepSWE. Same idea but they pulled the answers out, so the model has to actually figure it out. On DeepSWE, GPT-5.5 scored 70%. Gemini 3.5 Flash scored 28%. The two "tied" models weren't close at all. And that gap lines up with how different these tools actually feel to use. None of this makes benchmarks useless. They're fun to look at and they give you a rough starting point. But remember who makes most of them. A big score is a marketing asset. It's the number on the launch tweet. The keynote slide. The headline. So always take them with a grain of salt. What I actually do is I bounce between Opus and GPT all day. Not because one won a benchmark, but because I've built a feel for which one handles which kind of task. For serious work right now, those two are the only horses I really trust in this race. Building that feel isn't exciting. You take one task you actually need done, run it through different models/harnesses, and notice which one you trust with the result. Do that enough times and you stop reaching for the leaderboard. → A model that's perfect for someone else can be the wrong pick for you.