Back to the Drawing Board: Which AI Plays Best?
The Pi is finished and the course is written. So Kevin is starting something new: AI agents from different companies, sitting at the same table, under the same rules. Here’s the plan. Fair warning: it may take a while.
For twenty years the dream was one honest opponent that could tell you why it made its move. That dream is done. The table is fair (Part 2), the AI agent explains itself (Part 4), the Pi version is finished (Part 23), and the course is ready to give away (Part 25).
So, back to the drawing board. In Part 25 we said the table doesn’t care which AI you bring. Time to prove it.
The idea
In Part 22, students put six bots in one file and played a round robin where everyone got the same cards. This is the same game, with AI agents in the seats.
Kevin’s AI player runs on Claude. The plan is to clone it, give every clone exactly the same setup, and seat a different company’s AI behind each one. Same harness, same instructions, same tools, same answer format, same time and token limits. The only thing that changes is the model doing the thinking.
When everything else is equal, any difference at the table belongs to the AI. That turns a fun tournament into a fair test.
Three events
| Event | Format | What it shows |
|---|---|---|
| 1. The six-seat round robin | About six AIs from different companies, 600 hands, everyone dealt the same cards in turn | Who plays best at a full table |
| 2. The heads-up round robin | Every AI plays every other one-on-one, then the same cards again with seats swapped | Who beats whom, matchup by matchup |
| 3. The heads-up tournament | A knockout bracket: lose a match and you’re out | A champion, and how often the favorite really wins |
All three will be recorded the way every game on the site already is. You’ll be able to open any hand in Game Review, read each AI’s stated reason for every move, and download the data to check the results yourself.
Why Amazon Bedrock
Kevin’s AI player already runs on Amazon Bedrock, and Bedrock now offers models from many companies under one account: Anthropic, Amazon, Meta, Mistral, DeepSeek, OpenAI and others. That means one harness, one bill and one set of logs, with no pile of extra accounts and keys to protect. Companies that aren’t on Bedrock may get a seat in a later round.
The exact lineup isn’t picked yet. It depends on which models Kevin’s account can use, and what each one costs per hand.
More than chips
Winning chips is only one column. The others may be more interesting:
- Mistakes in the answer. How often does each AI break the answer format, so the safety net has to step in? That’s Part 5, measured company by company.
- Cost per hand. Some models cost many times more than others. Is the expensive one worth it?
- Speed. How long each one takes to decide.
- Reasoning. Does it think honestly about what the other player might hold, or does it slip into thinking it knows?
- Style. Who folds, who bluffs, who never lets go of a hand.
Making it fair
- Everyone gets the same cards. 600 hands is a small sample in poker, and luck can drown out skill. Dealing the same cards to every seat in turn is how you tell skill from luck.
- The same thinking time for all. Some AIs can be given extra time to reason and some can’t. Either nobody gets it, or everyone who can gets the same amount.
- Exact versions, with dates. AI models change quietly. The results will say which version played, and when.
- No crown from one run. If the margins are close, we’ll say so.
Why it may take a while
None of this is hard to imagine. All of it takes time to do right.
- The harness. One agent, with the model as a setting, tested on every AI for a few hands first.
- A small pilot. Fifty hands, to measure the real cost and catch problems before the bill gets big. Around 14,000 AI decisions go into the six-seat event alone.
- Heads-up first. It’s simpler, and the table already exists.
- Then the full table, then the bracket.
- Then the analysis and the write-up.
Kevin has a day job, and this is a hobby. It will happen in Saturdays, not sprints.
Make your prediction
Before a single hand is dealt: which company’s AI do you think wins? Does the most expensive model win? Does the one that “thinks” longest? Write your guess down, with a reason. When the results come in, check it, the same way you’d check any AI’s answer.
Source: Amazon Bedrock Model Catalog 2026 (an independent list, updated Oct 3, 2026). Which models an account can use varies by region and settings.
Written by Kevin Swinson with Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI. Describes plans as of October 24, 2026; nothing here has been played yet. No money is played, ever.