The $2/$10 Price War Quietly Changed Which Model I Reach For

Something odd happened in early October that most people scrolled past: three of the biggest AI labs landed on nearly the same price. OpenAI's GPT-6.1 Sol, Anthropic's Claude Sonnet 5.5, and Google's Gemini 4 Argon all settled around $2 per million input tokens and $10 per output, within about nine days of each other. Haiku 5.5 came in about 90% cheaper than the model it replaced, and GPT-6 Sol dropped roughly half. The price war is real, and it changes how I work.

For most of this year I treated the premium model as the default, because the gap felt worth it. The gap just narrowed to almost nothing on price while the quality gap stayed. So my habit of always reaching for the most expensive option stopped making sense, except on the specific tasks where a model is clearly better.

The numbers behind it are concrete. Gemini 4 Argon reports a 15% hallucination rate on one benchmark against 51% for GPT-6 Astra, which is why I point it at deep, document-heavy work. Claude's 5.5 family posts the strongest published agentic-coding scores, which is why my code goes there. The cheap Haiku handles the tagging and summarizing I used to waste a flagship on.

What I stopped doing is paying flagship prices for mechanical work. A route that sends the boring 80% to a fast small model and reserves the heavy model for judgment calls reads the same to the end user and costs a fraction. The dispatchers the labs are shipping point the same way.

The lesson isn't that cheap is best. It's that the price difference used to justify a default, and now it doesn't. Pick the model for the job, because the bill no longer makes the choice for you.