--- title: "The models were never the magic" url: "https://alex.zappa.dev/blog/models-were-never-the-magic" description: "Mathias Biilmann says the price of intelligence finally matters. I went one step further: the model stopped mattering entirely. The difference was always in the harness." --- Mathias Biilmann wrote this week that [the price of intelligence finally matters](https://biilmann.blog/articles/does-the-price-of-intelligence-matter/). The opener is the most honest thing a CEO has written about models all year: for two years he liked the idea of open weight models without wanting to use one, because cheap was their pitch and cheap was never his problem. Intelligence was. He even let DeepSeek V4 Flash typeset the article and Kimi K3 design the cover. On-message doesn’t cover it. “I’ve spent two years being right that price didn’t matter,” he says, “but I’m betting that won’t hold for another year.” He’s right about the split. One race stays at the frontier, where labs fight over state of the art and nobody shops on price. Another race makes intelligence ubiquitous and almost free. Open weights win the second one. But I’d push it further. Biilmann says DeepSeek V4 Flash is the first model that’s actually good enough and cheap enough to change which problems get AI pointed at them. I say the model stopped being the variable entirely. For most real work, they’re all the same. The difference was never the model. ## The sequel My last post ended on a spoiler: the browser and the IDE were only half the migration. [The AI agent itself](/blog/tool-migration-wavebox-jetbrains/) was a separate story, coming later. This is that story. June 2026 was the point of no return. Fable 5 launched fast on June 9, the US put export controls on it June 12, and Anthropic suspended access globally that same day. The next day I started testing every open model I could find, DeepSeek V4 included. Not out of enthusiasm. Arithmetic: today it’s Fable 5, tomorrow it’s every US model under export control, and my whole pipeline was standing on one of them. I already wrote about the slowdowns before the launch, the week-by-week crawl of Opus 4.8 that magically cleared up once Fable 5 shipped. What I didn’t write about is what I did after. I stopped renting my pipeline. A directive in Washington had just proved it could switch my delivery rhythm off in an afternoon. Fable 5 came back on July 1 once the controls lifted. The lesson didn’t. Whatever came next had to let me swap models without touching my workflow, and never again leave me holding one vendor’s mood. It’s not that I can’t work without them. LLMs are sugar: I’d survive cutting them out, but nobody should quit cold turkey on someone else’s schedule. Better to pick your dependency while the choice is still yours. ## The experiment So the harness changed. I run [Hermes](https://hermes-agent.nousresearch.com) as my main harness now, and the reason has nothing to do with any single model. It’s the plumbing: one harness, any model, tools and MCP servers wired once. That made the experiment easy to run. Local first. qwen3.6 for a while, then qwen3.8 27B when it dropped. Gemma for lighter jobs. Community fine-tunes too, including one where somebody literally trained a model and named it Fable 5. It works fine. The joke writes itself. Then the big open weights through APIs: DeepSeek V4 Flash, GLM 5.2, Kimi K2.7, which is fantastic for code, the fresh Kimi K3, Mistral, Devstral. Swap any of them mid-task and the work still gets done. Coding, PR reviews, changelogs, docs. All fine. All the same. ## It was never the model Here’s the part nobody selling you tokens wants to hear: Gemini, Grok, OpenAI, Claude. Point them all at the same task from the same harness and you get the same job done. I mean the same finished work, not the same tokens. The difference is in the harness. The tools the model can reach. The context you feed it. The memory, the retrieval, the MCP servers. A great model in a bare harness loses to a decent model with real context, every time. That stopped being a thesis months ago. It’s evidence now. One carve-out for research: Gemini and Grok are still the best tools for finding what just happened, tweets, Reddit threads, official sources, fresh stuff. Everything else is interchangeable for code and generic work. ## The Claude Code test I know people who won’t leave Claude Code because it’s the harness they trust. Fine. Keep it. But point the smartest open model at it and tell me what you lose. Install Ollama, take their subscription, connect a custom model from their servers into Claude Code. What happens: zero difference in coding quality, noticeably faster token generation, because their GPUs are faster than Anthropic’s. I have no idea how that economics works. I just watched it happen. The GDPR objection is the usual one. They host compute primarily in the US and route to Europe and Singapore for extra capacity, which settles the legal question for most teams. But honestly, if data residency is your real fear, stop sending your repo through someone else’s cloud at all. Run it local. That answer has been sitting there the whole time. ## The rebellion The open source mood has said the same thing for years. Be local. Own your software, own your hardware. Don’t be somebody’s subscription unit. It used to sound ideological. June 2026 made it engineering. I noticed the speed trick before I had the politics worked out. Opus 4.8 got slower every single week, then Fable 5 arrived fast and fresh, the same speed Opus had back when it was the new thing. A model you rent can slow down whenever the vendor needs you to notice the newer one. A model you run doesn’t negotiate. Biilmann is right about the split, and I’ll add the part he left out: the ubiquity race eats the middle of his chart faster than the labs expect, because the models under the frontier are already good enough. Europe especially lost the luxury of choosing. When one afternoon in Washington can switch off your pipeline, open weights stop being an ideology and become the audit trail. The models were never the magic. The harness was. Own it. May the 4th be with you, Alex - [PreviousI walked away from WaveBox and JetBrains in the same month](/blog/tool-migration-wavebox-jetbrains/)