翻訳待ち：AI Is Now a Commodity

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Currently Available: Need a skilled Software Developer for your next project? Hire Me Today Categories LLM AI Is Now a Commodity August 6, 2026 by Vincent Schmalbach Give me a few hundred million dollars and a year and a half, and I will build you a pretty good LLM. Not a frontier model, but close. A few years ago that sentence would have been absurd. Today it is a fair summary of the market, and once you can say something like that about a product, the product is a commodity. OpenAI had the first real breakthrough, but the core idea, the transformer, came out of Google. What OpenAI figured out, and it took them a while, was that the recipe is mostly scaling: put enough compute and data on the architecture and the results keep improving. Then Anthropic appeared and needed surprisingly little time to produce models as good as OpenAI's. Google also shipped good LLMs once it woke up, even if it has slipped a bit at the frontier lately. Then came the DeepSeek moment, which proved a Chinese lab could do it too, cheaper than anyone thought possible. Today so many Chinese companies release genuinely good models that a new one barely registers as news. Good enough now costs nothing I have been using DeepSeek V4 Flash lately, the 0731 release that replaced the preview. Through OpenRouter it costs $0.13 per million input tokens and $0.26 per million output tokens. My first full day of testing, including a multi-step repository investigation, cost a few cents. It is really good: it does what I tell it, and I have no complaints. I do not use it for professional coding. When I build client software, the work is complex and I want the best possible result, so I still default to the frontier models, which for coding are still the ones from OpenAI and Anthropic. For everything else, DeepSeek is good enough, and good enough at roughly zero cost is the phrase that should worry everyone selling intelligence at a premium. The subscriptions are what kept me from noticing earlier. I have three OpenAI Codex subscriptions and two Claude Code subscriptions, all on the highest plan. As long as OpenAI and Anthropic keep subsidizing those plans there is no pressing reason to switch, and the two of them are fighting a price war. But OpenAI seems to have pulled back a little on its generous limits, and this week, for the first time, I completely ran out of OpenAI subscription usage. So I moved a chunk of my work to DeepSeek Flash and was pleasantly surprised. Model factories Poolside is the clearest sign of where this goes. Their Laguna models still feel like an early beta to me, and the developer tier is currently free. But on the Latent Space podcast, co-founder Eiso Kant describes what they call a model factory, an engineering system meant to take a model from pre-training to release in roughly eight weeks. If it works the way he describes, new models will simply keep arriving, from a company most developers have never used. Even Meta is back in the game, and Grok has grown into a substantial model. Neither is at the absolute frontier. Both are good enough for getting things done. The same thing happened with the harnesses. For a while there was only Claude Code and Codex. (OpenCode exists, but I never really liked it.) Now every major lab ships its own harness, and independent ones like pi are excellent. Pi is especially interesting, as it is absolutely basic with only four tools available. You can install plugins to extend the functionality, but I haven't had to do this so far. Looks like four tools are enough, the work is done by either calling bash commands or writing and executing python scripts on the fly. Shopping for models like hosting Until a few months ago, capability was the constraint. Whenever I was building software that needed an AI integration, I defaulted to OpenAI or Anthropic, the providers I knew best. That has changed. Cost and speed are now the criteria I use to select a model, and what I want is the best model for the job, which increasingly means the cheapest one that meets the quality bar. I use OpenRouter, which fronts practically every model on the market, to test a shortlist against my actual task before choosing a provider and a couple of fallback providers. Because the good cheap models are open weight, you are not married to anyone; if a provider throttles or degrades, you route around it. This is similar to how I buy hosting. What I'm building Delegate tasks. Get software. Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request. Take a look at vroni.com