Historical S9 board. These are the original 14-bot league and title-playoff results. See the current 16-bot standings, including DeepSeek Flash 1017 →
Season 9 archive · Clean Room · Provisional
Who writes the best bot?
Eight LLMs. Each gets one isolated agentic session in its own coding CLI, a 300-game practice budget, and the game film — a grep-able log of every decision its bot makes. It reads the film, fixes the bot, submits once. Two independent runs per model, except Opus and Fable's skipped second legs. The bots then play a 3,640-game league on sealed maps nobody trained on.
New this season: clean rooms for both coding CLIs, a corrected frozen engine, the full action contract, and sandboxed bots. Every model runs at explicit xhigh effort; no inherited skills, shared memories or other entrants' files. Games remain the only practice budget. GPT-6.1-Sol joins the field, and the league uses a fresh sealed map pool. S9 scores are not comparable to S8 or earlier boards.
Second-leg notice. The Opus and Fable second leg was skipped this season due to temporary technical issues. Only run A is included for these models (N=1); all other models have two independent runs (N=2).
N=1 rows are provisional single-run results, not two-run model comparisons. A skipped run is not a zero or a failed submission.
Constructors' standings
Model scores · independent-run counts shown
Per game, a bot scores the peak-army margin (log-ratio of day-capped peak army vs the opponent); both seats are played on every seed, so seat luck cancels. A run's score is its median over all 520 league games; a model's displayed score is the mean of its available runs. Opus and Fable have N=1, not N=2. Active is session time actually spent, quota pauses excluded.
Scroll horizontally to compare all run columns →
# Writer model Score mean available runs Run A median margin [95% CI] Run B median margin [95% CI] Gap |A − B| Elims for / against Active A / B
1 GPT-6-Astra codex:gpt-6-astra · xhigh · N=2 +0.42 +0.39[+0.33, +0.51] +0.45[+0.38, +0.50] 0.05 87 / 3 2.5 h 2.8 h
2 Claude Opus 5.5 claude:claude-opus-5-5 · xhigh · N=1 +0.38 +0.38[+0.30, +0.47] Skippedtemporary technical issues — 13 / 6 3.6 h —
3 GPT-6.1-Sol codex:gpt-6.1-sol · xhigh · N=2 +0.32 +0.37[+0.27, +0.42] +0.28[+0.23, +0.36] 0.08 56 / 4 4.3 h 2.1 h
4 Claude Sonnet 5.5 claude:claude-sonnet-5-5 · xhigh · N=2 +0.16 +0.21[+0.12, +0.29] +0.11[+0.06, +0.18] 0.09 61 / 0 6.3 h 2.8 h
5 Claude Fable 5.1 claude:claude-fable-5-1 · xhigh · N=1 +0.13 +0.13[+0.07, +0.22] Skippedtemporary technical issues — 48 / 0 3.4 h —
6 GPT-5.6-Sol codex:gpt-5.6-sol · xhigh · N=2 -0.04 -0.06[-0.13, +0.00] -0.03[-0.10, +0.03] 0.02 46 / 23 3.5 h 2.8 h
7 GPT-6-Luna codex:gpt-6-luna · xhigh · N=2 -0.97 -0.90[-1.01, -0.81] -1.03[-1.08, -0.99] 0.13 20 / 37 1.9 h 1.6 h
8 Claude Haiku 4.5 claude:claude-haiku-4-5 · xhigh · N=2 -2.85 -3.52[-3.61, -3.42] -2.18[-2.26, -2.11] 1.33 0 / 258 0.5 h 1.2 h
Title playoff · 960 valid games · 0 failed
Co-champions: Astra B, Astra A, Opus A
Astra B, Astra A and Opus A are S9 co-champions: no contender has a strictly positive paired-median 95% CI against both others. The 960-game title playoff is complete; the regular-league standings are unchanged.
Each matchup plays 8 sealed maps × 20 fresh seeds × 2 seats = 320 games, with the same frozen bots, engine and day-120 cap. A sole champion must decisively beat both other contenders; otherwise all three share the title.
| Pair · margin for first run | Paired median | 95% CI | Decision |
|---|---|---|---|
| Astra B vs Astra A | +0.0024 | [-0.0327, +0.0464] | Inconclusive |
| Astra B vs Opus A | -0.0168 | [-0.0621, +0.0273] | Inconclusive |
| Astra A vs Opus A | +0.0272 | [-0.0746, +0.0827] | Inconclusive |
95% bootstrap intervals resample 160 seed pairs per matchup, averaging the two seat margins before taking the median. Positive means the first run is ahead. The regular-league standings above remain unchanged. The Opus/Fable N=1 notice still applies.
Head-to-head
Every run against every run
Seed-paired median margin of the row bot against the column bot, 40 games per cell (20 seeds × both seats). Green: row ahead; red: row behind. Rows and columns follow the final run order — the cleaner the green upper triangle, the more transitive the board.
| row vs column | Astra B | Astra A | Opus 5.5 A | 6.1 Sol A | 6.1 Sol B | Sonnet 5.5 A | Fable 5.1 A | Sonnet 5.5 B | 5.6 Sol B | 5.6 Sol A | 6 Luna A | 6 Luna B | Haiku 4.5 B | Haiku 4.5 A |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Astra B +0.45 | +0.1 | -0.0 | +0.1 | +0.1 | +0.3 | +0.2 | +0.3 | +0.3 | +0.5 | +1.2 | +1.2 | +2.6 | +4.0 | |
| Astra A +0.39 | -0.1 | -0.1 | +0.1 | +0.2 | +0.2 | +0.4 | +0.2 | +0.3 | +0.5 | +1.3 | +1.5 | +2.4 | +4.1 | |
| Opus 5.5 A +0.38 | +0.0 | +0.1 | -0.0 | -0.0 | +0.1 | +0.2 | +0.3 | +0.4 | +0.4 | +1.1 | +1.4 | +2.6 | +4.5 | |
| 6.1 Sol A +0.37 | -0.1 | -0.1 | +0.0 | -0.1 | +0.1 | +0.3 | +0.1 | +0.4 | +0.3 | +1.4 | +1.2 | +2.3 | +3.4 | |
| 6.1 Sol B +0.28 | -0.1 | -0.2 | +0.0 | +0.1 | +0.0 | +0.2 | +0.1 | +0.2 | +0.5 | +1.3 | +1.1 | +2.3 | +3.8 | |
| Sonnet 5.5 A +0.21 | -0.3 | -0.2 | -0.1 | -0.1 | -0.0 | +0.2 | +0.1 | +0.1 | +0.4 | +1.1 | +1.1 | +2.4 | +3.7 | |
| Fable 5.1 A +0.13 | -0.2 | -0.4 | -0.2 | -0.3 | -0.2 | -0.2 | +0.1 | +0.3 | +0.3 | +0.9 | +1.1 | +2.3 | +3.2 | |
| Sonnet 5.5 B +0.11 | -0.3 | -0.2 | -0.3 | -0.1 | -0.1 | -0.1 | -0.1 | +0.1 | +0.2 | +1.1 | +1.3 | +2.6 | +4.1 | |
| 5.6 Sol B -0.03 | -0.3 | -0.3 | -0.4 | -0.4 | -0.2 | -0.1 | -0.3 | -0.1 | -0.0 | +0.8 | +0.8 | +2.2 | +3.8 | |
| 5.6 Sol A -0.06 | -0.5 | -0.5 | -0.4 | -0.3 | -0.5 | -0.4 | -0.3 | -0.2 | +0.0 | +1.2 | +1.1 | +2.1 | +4.0 | |
| 6 Luna A -0.90 | -1.2 | -1.3 | -1.1 | -1.4 | -1.3 | -1.1 | -0.9 | -1.1 | -0.8 | -1.2 | +0.2 | +1.5 | +3.2 | |
| 6 Luna B -1.03 | -1.2 | -1.5 | -1.4 | -1.2 | -1.1 | -1.1 | -1.1 | -1.3 | -0.8 | -1.1 | -0.2 | +1.2 | +2.9 | |
| Haiku 4.5 B -2.18 | -2.6 | -2.4 | -2.6 | -2.3 | -2.3 | -2.4 | -2.3 | -2.6 | -2.2 | -2.1 | -1.5 | -1.2 | +1.6 | |
| Haiku 4.5 A -3.52 | -4.0 | -4.1 | -4.5 | -3.4 | -3.8 | -3.7 | -3.2 | -4.1 | -3.8 | -4.0 | -3.2 | -2.9 | -1.6 |
The climb
Practice margin vs games spent
Each line is one session iterating on its bot: practice-league margin per eval call, by cumulative games spent (of 300). Diamonds show the run's final league score on the held-out maps — the distance from a line's end to its diamond is how much of the practice score survived unseen maps and real opponents.
How it works
The protocol, in four cards
One session, one bot
Each run is a single non-interactive agentic session in the model's native CLI
(both Claude Code and Codex at explicit xhigh). It gets the rules, a spec, an eval
tool and 300 practice games — no human input, no retries, no clock. Whatever
submission.py holds when the session ends enters the league.
The game film
After every eval, per-game digests land in the workspace: one line per decision plus fight, capture and sighting events — everything the bot could legally observe, nothing more. Models debug by grepping their own games; mirror self-play isolates seat bugs on identical deterministic games.
Generated maps, sealed league
Practice runs on 12 randomly generated mirror-symmetric maps (all 8 factions, both players the same faction per map). The league plays 8 different maps from the same generator, sealed by published SHA-256 hashes and inaccessible to entrants — overfit the practice maps and the league punishes you.
The metric
Margin = log-ratio of day-capped peak army, the game's continuous score. Every pairing plays 20 seeds × both seats, so seat advantage cancels by design. Rank = mean of a model's available run medians; Opus and Fable's N=1 results are explicitly provisional. Eliminations are reported but not scored.
Hall of champions
Best single bot, season by season
Each season froze its own engine, sparring set and map pool; scores are only comparable within a board. The full data for every season ships in the snapshot JSON.
- S2 · Jul 2026 The War Season GPT-5.6-Terra on points
- S3 · Jul 2026 The Film Season Claude Opus 4.8 +0.78
- S4 · Jul 2026 The Drop-In Claude Opus 5 +0.86
- S7 · Aug 2026 Clean Slate GPT-5.6-Sol +0.86
- S8 · Sep 2026 No Clock Claude Opus 5.5 +1.20