In April I tested Claude Code in embedded live: I sat it down next to Codex in front of the new NUCLEO-C562RE<\/a> board, gave them a BMP280<\/a> sensor, the brand-new HAL2 and very general instructions. Claude Code hit its usage limit halfway through. Codex finished the simplest version, and after two hours temperature and pressure were streaming to the terminal, but it also burned through almost its entire limit. I described it in my May article and closed with the sentence that AI is a great assistant, but not yet a replacement.<\/p>\n\n\n\n More than five months have passed since that live stream. Today half of that article is out of date. And that is what this article is about: statements about AI become outdated faster than anything I know from embedded<\/strong>. Mine included.<\/p>\n\n\n\n Let me start here, because some people treat that live stream as my first contact with AI. What you see in public is not everything.<\/p>\n\n\n\n I have been working with AI for a long time, programming included. I have been programming STM32 and other microcontrollers with its help for quite a while. In March 2025 I gave a talk at a meetup titled \u201cHow AI lets me program things I have no idea about\u201d. And back then I really wasn’t fully happy with how GPT or Claude handled embedded in chat form.<\/p>\n\n\n\n Literally while I was preparing that talk, Claude Code was announced. It was a novelty, and today, from what I see, it is the most important AI agent!<\/p>\n\n\n\n My claims that LLMs handle embedded only so-so were not made up. Nor did they come from fear for my programming course sales (because that’s the easiest thing to accuse me of ;)). They came from practice and programming experience.<\/strong><\/p>\n\n\n\n So during the April live I wasn’t a sceptic looking for proof. The setup was simple: complete freedom for the agent and a minimal prompt written on the fly<\/strong>. Total free rein, no guidance at all. I wanted to see whether the agent would really do the job for me, as commenters under my ads claimed.<\/p>\n\n\n\n I had been working with chat for a long time, but I was only starting out with agents, and I said so openly on the live<\/strong>. Whoever uses my words from that stream today probably didn’t watch it very carefully \ud83d\ude42<\/p>\n\n\n\n And what came out of that free rein worried me. An agent that got no project, just a single instruction, did whatever it thought was right.<\/p>\n\n\n\n Since the model I was using back then, Anthropic has released four more Opus models: 4.7 still in April, 4.8 in May, Opus 5 in July and Opus 5.5 in September. On the other side, GPT-6 Sol appeared on 22 September, and on 29 September it was replaced by 6.1 Sol. That model lived for just one week!<\/strong> How are you supposed to record anything about AI when it changes this fast? \ud83d\ude05<\/p>\n\n\n\n Anthropic and OpenAI are really fighting hard for dominance. They release newer and newer models at shorter and shorter intervals, and their capabilities grow exponentially. This is measured by METR<\/a>, an independent organisation that studies what models can do on their own. It checks how long a programming task a model can complete in half of its attempts, with length measured as the time an experienced human would need for it. According to its January measurements<\/a>, that length doubles roughly every four months<\/strong>.<\/p>\n\n\n\n The human brain struggles with that kind of progress. We think linearly: if it was weak six months ago, today it’s a bit less weak. It isn’t.<\/p>\n\n\n\n That’s why I’m not repeating that experiment. The result would be outdated before I managed to write it up. Pointless.<\/p>\n\n\n\n A few numbers, all as of 2 October 2026:<\/p>\n\n\n\n The two cost figures come from the same week and say different things, because they refer to different settings. Maximum effort is not a mode for everyday work, and that’s exactly how I worked in April: Opus 4.6 on xhigh.<\/p>\n\n\n\n I don’t follow every piece of news myself, and I don’t intend to. So I simplified model choice for myself: I use the newest one, at default effort<\/strong>. For a specific task I move it one notch up or down at most. That’s it.<\/p>\n\n\n\n Price alone says little, so let me describe one benchmark in more detail.<\/p>\n\n\n\n Terminal-Bench<\/a> is an open benchmark developed by the community and the Harbor team. The agent gets a terminal in a container and a task to complete end to end: compile, configure, fix, run. It doesn’t write a single function, it has to deliver a working result. When it’s done, a separate script checks the final state. A task only counts as passed if all tests pass. There is no partial credit.<\/p>\n\n\n\n When I did the live stream, version 2.0 with 89 tasks was current. According to results published by Anthropic, Opus 4.6 scored about 65% on it, and Opus 4.7, released just before the live, about 69% (overview<\/a>). On the revised version 2.1, Opus 4.8 reached 79% in May and GPT-5.5 83% (leaderboard<\/a>).<\/p>\n\n\n\n The benchmark stopped telling models apart, so version 4.0 was created: 66 completely new tasks from programming, science, machine learning, hardware and security. Half of them take an expert at least four hours, the longest ones 60 hours. The agent gets eight hours for each.<\/p>\n\n\n\nIt wasn’t my first attempt<\/h2>\n\n\n\n
What aged: models and cost<\/h2>\n\n\n\n
\n
What it means for programming: Terminal-Bench<\/h3>\n\n\n\n