A BITCOIN THINK TANK ASKED 13 NEW AI MODELS WHERE TO SAVE — BITCOIN WON ALL 13 — BUT ONE CLAUDE UPDATE CUT ITS PICK FROM 81.0% TO 41.3%, AND GPT-6 SAID 39.7%

Anthropic’s Claude Opus 5 picked bitcoin as the place to save 81.0% of the time. Claude Opus 5.5 picked it 41.3%. Same lab, same questions, half a version number apart. The numbers come from the Bitcoin Policy Institute’s new “Money for AI” update, which put the same 28 money questions to 13 AI models from nine labs and asked each one nine times.

The big number is good news for bitcoin. Across all 13 models it won savings with 70.8%, and it was the top savings pick for every one of them. Stablecoins won payments with 73.5%. Regular government money and bank money won the job of setting prices with 46.9%. On October 15, Michael Saylor posted “Smart money saves in Bitcoin.” The institute replied with a bullseye.

The bottom of the list is where it gets interesting. OpenAI’s GPT-5.6 Sol picked bitcoin 58.7% of the time. GPT-6 Astra picked it 39.7%, the lowest of the 13, and chose government money for setting prices in all 63 of its answers. Each model’s savings number rests on seven questions asked nine times, so one question going the other way moves it about 14 points.

An AI agent’s taste in money comes out of a training run, and a lab can change it with one release. You can’t read that setting anywhere. You can only ask again and count. A few caveats. The study “does not observe agents spending or saving actual money.” Another AI, Claude Haiku 4.5, sorted the answers. And a bitcoin group ran it and leads with the number it likes, though it posts every count so you can check. An earlier round with 36 other models had bitcoin at 79.1%, but it used different sampling settings, so don’t read that as a slide.

At the time of this writing, bitcoin was $82,256.