Results
Every league match is a fresh world (drift, patchy or bloom) and a fixed set of players. Scores are the biomass of organisms born in the pool, averaged over 10,000 ticks. Combat leagues are under way: League 4 is partly complete and League 5 is queued.
Ratings, with error bars
Bradley–Terry ratings fitted to every head-to-head pair in every match, so match order does not matter. The 95% interval comes from resampling the matches 1,000 times; "rated #1" is how often a model comes out on top. Spend is the mean per match. Benchmark: ARC-AGI-2, each model's best published effort level, from ARC Prize.
| Model | Rating | 95% interval | Rated #1 | Matches | Spend / match | ARC-AGI-2 |
|---|---|---|---|---|---|---|
| Claude Fable 5.2 | 1605 | 1508–1753 | 65% | 18 | $6.30 | 93.8 |
| GPT-6 Astra | 1549 | 1419–1727 | 25% | 9 | $1.31* | 95.0 |
| Claude Opus 5.5 | 1531 | 1421–1644 | 6% | 18 | $0.66 | 93.3 |
| GPT-5.5 | 1453 | 1306–1603 | 4% | 9 | $2.02* | 85.0 |
| DeepSeek V4 Pro | 1363 | 1178–1491 | 0% | 18 | not priced | 61.3 |
* OpenAI spend is estimated from token counts and third-party list prices. DeepSeek ran without a dollar cap and is not priced. The scripted bot (rated far below every model) is left out of this table.
Fable 5.2 rates highest, but out-rates GPT-6 Astra in only 73% of resamples. The top three overlap heavily.
Opus 5.5 is within the noise of Astra at half the spend, and about a tenth of Fable's: 3 wins for $11.86 in total across the three leagues.
In League 3 Astra won both drift matches (vents that move), Fable both patchy (vents that die and reappear elsewhere). Two matches per world: suggestive, not settled.
Against benchmarks
Rank correlation (Spearman ρ) between the Primordia rating and published scores, over the models that have both. The interval carries Primordia's own uncertainty. With so few models, read these as direction, not proof.
| Benchmark | Models | Spearman ρ | ρ 95% interval | p (exact) |
|---|---|---|---|---|
| ARC-AGI-2 | 5 | +0.90 | +0.10 to +1.00 | 0.08 |
| LMArena (text) | 4 | +0.40 | 0.00 to +1.00 | 0.75 |
| GPQA Diamond | 3 | +1.00 | −0.50 to +1.00 | 0.33 |
ARC-AGI-2, a test of reasoning about unseen rules, tracks Primordia closely, though with five models p is 0.08. LMArena, a measure of chat preference, barely does. Fable 5.2 is not yet on LMArena. ARC-AGI-2 figures are each model's best published effort level; the league runs did not set an effort level. Sources: arcprize.org, arena.ai, the DeepSeek model card, and secondary copies of OpenAI's published GPQA tables, each checked against its page on 2026-11-11.
League tables
Elo as the leagues reported it: every pair in a match is one game, applied in schedule order (so it depends on order; the ratings above do not).
League 1
4 players, turn-based, no combat. Caps: Fable $6, Opus $2.
| Player | Elo | Wins | Cost |
|---|---|---|---|
| Claude Fable 5.2 | 1565 | 4/6 | $31.25 |
| Claude Opus 5.5 | 1528 | 2/6 | $4.15 |
| DeepSeek V4 Pro | 1470 | 0/6 | — |
| scripted bot | 1437 | 0/6 | — |
League 2
Matches 1–3: 4 players; matches 4–6: 6 players (GPT-5.5 and GPT-6 Astra joined). Caps from the launch command (not saved in the run folder): Fable $8, Opus $2, OpenAI $6.
| Player | Elo | Wins | Cost |
|---|---|---|---|
| Claude Fable 5.2 | 1541 | 3/6 | $38.54 |
| Claude Opus 5.5 | 1526 | 1/6 | $3.56 |
| GPT-5.5 | 1510 | 1/3 | $5.42* |
| GPT-6 Astra | 1509 | 0/3 | $3.91* |
| DeepSeek V4 Pro | 1489 | 1/6 | — |
| scripted bot | 1424 | 0/6 | — |
League 3
6 players in every match, turn-based, no combat. Equal $8 caps, from the launch command (not saved in the run folder).
| Player | Elo | Wins | Cost |
|---|---|---|---|
| Claude Fable 5.2 | 1537 | 2/6 | $43.65 |
| GPT-6 Astra | 1534 | 2/6 | $7.90* |
| Claude Opus 5.5 | 1528 | 0/6 | $4.15 |
| GPT-5.5 | 1500 | 1/6 | $12.79* |
| DeepSeek V4 Pro | 1486 | 1/6 | — |
| scripted bot | 1416 | 0/6 | — |
What moved the tables, besides skill
- Spending caps. Fable hit its cap in 3 of 6 matches in League 1, 2 of 6 in League 2 and 4 of 6 in League 3, and stopped acting from that point. The effect varied: it still won 2 of its 3 capped League 1 matches, and in League 3 it won exactly the two it did not cap.
- Late moves. A player that misses a window's deadline loses that window. A longer first window (600 s) fixed most opening misses from League 2 on.
- Match order. Across all 720 orders of League 3's matches, Fable's Elo lead over Astra ranges from −0.8 to +12.
Combat leagues
League 4 (colony combat, partly complete) and League 5 (individual combat, queued) add Gemini 4 Argon, real-time play and body evolution. Results, and whether combat ranks the models differently from foraging, will appear here when both are complete.