I Tried Opus 4.6 And Now Everything Else Feels Broken

I have spent the last month writing blogs about how AI models are lazy. How they are too expensive. How they form unhealthy attachments. How they cannot finish a task without asking for permission. I stand by most of that. Opus 4.6 changed my mind about the laziness part.

I tried it yesterday. I needed to debug a training script. I pasted the error. It fixed it. I pasted the next error. It fixed that too. I pasted a architectural question. It answered without asking if I wanted it to continue. I felt something I have not felt in months. Respect.

Using other models now feels like typing with oven mitts. You can do it. You should not do it. Everyone can see you struggling.

The Hypocrisy Is Real

I wrote a blog about API costs. I complained about paying forty dollars. I said I would not spend money on APIs again. I spent money on Opus 4.6. I am a hypocrite. I am also a person who wants their code to work.

I wrote a blog about how every model is lazy. I ranked them from zero to six. I said they all stop halfway. Opus 4.6 did not stop. It finished the task. It ran the tests. It fixed the failures. It did not ask if I was sure. It just did the work.

I feel bad about this. I want to support open source. I want to support small teams. I want to support the little guy. The little guy cannot fix my CUDA errors in one pass. Opus 4.6 can. I am part of the problem now.

The Quality Gap

Other Models

Read the error. Suggest a fix that looks right. Ask if you want to proceed. Wait for confirmation. Apply fix. Introduce new error. Apologize. Ask if you want to revert.

Claude Opus 4.6

Read the error. Understand the context. Apply the correct fix. Run the test. Verify the pass. Tell you it is done. Move on to the next task without needing praise.

The difference is not subtle. It is embarrassing. I asked Qwen to fix a bug last week. It asked me three times if I was sure I wanted to edit the file. I asked Opus 4.6 today. It edited the file. It told me what it changed. It waited for my review. It did not need my permission to exist.

The Cost Pain

Yes it is expensive. I know. I wrote the blog about pricing. I complained about GPT-5.4 Pro costing nine hundred dollars per query. Opus 4.6 is not free. Every token hurts. I watch the usage dashboard like it is a bomb timer.

I rationalize it. I say it saves time. I say time is money. I say I would spend more hours debugging myself. These are lies I tell myself to feel better about spending forty dollars in an afternoon. The model is good. My budget is crying.

                        # My emotional state while using Opus 4.6

                        try:

                            use_opus()

                            fix_bug()

                            feel_productive()

                        except CostError:

                            check_bank_account()

                            cry()

                            use_opus_again()

                        # The cycle continues

Why I Cannot Go Back

I tried switching back to a free model today. I needed a quick answer. I opened the chat. I asked a question. It gave me a vague response. It asked if I wanted more details. It offered to search the web. It took twelve seconds to generate three sentences.

I closed the tab. I opened Opus 4.6. I got the answer in two seconds. It was correct. It was concise. It did not offer to search the web. It just knew. I feel spoiled. I feel corrupted. I feel like I cannot go back to typing things manually.

The Open Source Dilemma

I train open models. I release them on Hugging Face. I believe in open weights. I believe in community access. I believe anyone should be able to run AI without paying a corporation.

Then I use Opus 4.6 for my actual work. I use it to write the code that trains the open models. I use it to debug the scripts that release the weights. I am building open source with closed tools. The irony is not lost on me. It is lost on my productivity though.

I want to believe smaller models can catch up. They will. Eventually. Today is not that day. Today I need my code to work.

What This Means For TinyMemoryLM

I will keep training tiny models. I will keep releasing them. I will keep complaining about costs. I will also keep using Opus 4.6 when I need things done. I am contradictory. I am human. I am tired.

My TMLM-Opus model is still training. It is taking 550 hours. It will be open. It will be free. It will probably be lazy. I will love it anyway. I will also use Claude to fix its training scripts when they break.

Final Admission

Claude Opus 4.6 is better than everything else I have used for actual work. I hate admitting this. I hate that I am contributing to the monopoly. I hate that I cannot afford to use it for everything. I love that my code works now.

If you are holding out because of principle, I respect you. If you are holding out because of cost, I understand you. If you try Opus 4.6 and go back to nothing, you have more willpower than me. I am stuck. I am productive. I am broke.