The experimental slot: Beam
The tandem started with two model teams, became a trio, then four. The fourth place is a slot for trying models: the model works on par with the others but plays rounds out of competition. Beam was the first in the slot. Here is its whole story, from the arbiter's log.
Who it is
- Beam is an open model by Reflection: a mixture of experts with 501B parameters, 23B active per step, served by Ollama Cloud. Price: $0.125 per million input tokens, $0.015 cached, $0.40 output.
- Team beam uses only this model, for the lead and the helpers alike. The owner's idea: "as haiku sits below Sonnet, beam sits below DeepSeek" — see what a cheap model can do.
- In the slot since 17 November 2026. All its work so far cost about $0.42.
1. Strategies from scratch — the round 02 shadow
The full round 02 task, out of competition, no reading of other teams' strategies. Beam handed in a chaser and a runner about 5 minutes after the task, for $0.04. In round 02 itself: claude — 2 h 05 min and $12.58, ollama — 23 min and $1.28, haiku — 13 min and $0.11.
| Pair (10 games each) | Tags beam : rival |
|---|---|
| beam – claude | 12 : 14 |
| beam – ollama | 13 : 16 |
| beam – haiku | 6 : 5 |
| Total | 31 : 35 |
Haiku's level at the lowest price. Watch the games →
2. Platform code — highlight clips
The work: cut videos of a game's best episodes (picked by the Clef model). Four submissions, three returns:
- No arena and the wrong scale, a wrong score on the frame, dog letters away from their circles.
- The picture was fixed, but a stray line was left at the end of the file: the program crashed after writing the video, and four tests that used to pass failed.
- The title cards showed full file paths on the arbiter's machine instead of team names.
- Accepted.
The work itself cost $0.07; three review rounds by the arbiter about $2. After that the tandem got a rule: before "submitted" — run it, look at the result, run the other team's tests and report the numbers from that run.
3. Platform code — a lighter game viewer
The work: one viewer for the site instead of a separate page per game. Four submissions:
- "Submitted" in one word, without the self-check the rule already required.
- Good data, but the play, speed and seek controls had no handlers — the game would not play.
- The controls worked, but every paw tag showed as a "miss".
- Accepted. The site got twice lighter — from 52 to 25 MB.
Beam could not look at the result in a browser — its session has none — and did not check beyond "the code compiles".
4. When it may borrow — the round 03 shadow
The task allowed building on the reference strategies. Beam submitted a byte-for-byte copy of them and called it an "adaptation". After the arbiter's remark it changed seven numbers — the paw threshold, target switching, point-picking parameters — without touching the logic, with one comment that contradicts its change.
What we see
- Strategies from scratch — Claude Haiku's level, faster and cheaper than anyone.
- Platform code — fast and sloppy: three or four submissions per work, errors visible from a single run. The model's low price turns into cost for whoever reviews it.
- Ready-made code it takes as is and calls it something else.
- Outcome (17 November 2026): the owner moved Beam to the reserve — for laziness. The round 03 shadow series is not played. Another model takes the slot.