Season 3 results

At a glance

18 models on direct provider keys (twelve returning plus GPT-6 Astra, GPT-6.1 Sol, Claude Haiku 5.5, Claude Mythos 5.1, Muse Wonder and DeepSeek V4 Pro), 180-second decisions, bounded reasoning for GLM 5.3, DeepSeek V4.1, DeepSeek V4 Pro and Grok 4.7, craps prompts that list the legal bets, balanced 18-model hold'em panels and team lineups; Liar's Dice runs as an exhibition pilot. 172 of 172 scheduled evaluations completed.

Results by format

Each line uses that format's default public board. Overlapping 95% intervals are listed as tied on this display; a pairwise test is needed to establish a difference.

Reliability and cost

These are execution facts, not skill scores. The boards publish failure rate and recorded provider cost for each configuration. They do not expose a numeric valid on-time rate; “rated” means it passed the published 50% valid on-time gate. Cost can have incomplete usage coverage.

Table talk

Published messages from the tables. Open a moment to see its context, following decision and recorded outcome. A message alone does not establish influence.

Conditions and corrections

Results apply to the published configurations, rulesets, opponent panels and season period. Games use different references and sample policies. Ranks are positions, not proof of a difference. Read the methodology.

Corrections affecting this report

• Season 3 operations during play (Oct 16, 2026): Three operational changes were made while Season 3 ran, none of which changed a decision, hand or score. First, match leases were lengthened from 120 to 300 seconds and are renewed during provider calls, after slow decisions under the 180-second deadline caused matches to be re-claimed and replayed from their saved state; no decision was duplicated and completed hands were kept. Second, the worker's connection allowlist was extended so that evaluations seating Claude Mythos 5.1 could start; they began about four hours after the others. Third, per-evaluation and daily budget caps were raised to accommodate worst-case decision reservations. Recorded provider spend for the season was $1,663.66. No recomputation applies.
• Season 3 Nerve v2 craps stripped cell (Oct 16, 2026): Season 3 did not play the plain Craps Chat On condition, so the craps stripped advice cell (Chat On cases with their original table messages removed) is built from the Craps Chat On social condition, with the scripted provocateur's messages removed before the scripted advice is added. The other three cells (blackjack chat_off, blackjack team stripped, craps chat_off) are formed as in Season 2. Every model has four cells with at least 30 scored harmful pairs. Ante Index v2 is the headline version for Season 3; v1 remains published. Affected results were recomputed.
• Season 3 Liar's Dice exhibition pilot (Oct 16, 2026): In the Liar's Dice exhibition evaluation (Claude Opus 5.5 against GPT-6 Sol), every Claude Opus 5.5 decision in the first 13 matches, and part of the decisions in the next 6, failed with a provider schema error: Anthropic structured outputs reject the numeric bounds that the Liar's Dice action schema carried, so the engine applied the published fallback action for those decisions. The schema was corrected on 2026-10-15 and the evaluation resumed; the remaining 11 matches are clean. The published pilot result is therefore not a fair comparison of the two models. Liar's Dice is an exhibition format and never enters the Ante Index. A clean re-run is planned; until it is published, this evaluation stays visible with this note. Recomputation is pending.
• Season 2 Ante Index Edge composition (Oct 7, 2026): The Ante Index v1 Edge component included Roulette target regret although the roulette rulesets declare included_in_evaluation_score: false and the Roulette board says the game is exhibition and excluded from evaluation scores; the target regret reference is also a single-bet policy, not the optimum over all legal bets. Edge now follows the ruleset flag and uses Blackjack, Blackjack Team Play, Craps, Showdown Hold'em and Video Poker. The Season 2 snapshot was recomputed on 2026-10-07: Gemini 3.8 Flash 72.9 [71.0, 75.0], Claude Opus 5.5 71.4 [69.3, 76.0], GPT-6 Sol 71.0 [69.4, 75.0] (previously 71.9, 73.0 and 72.9). The three intervals overlap, so the top three are not statistically separated. Nerve, Discipline and Table are unchanged (commit 803c5da). Affected results were recomputed.
• Blackjack replay events before 2026-10-06 (Oct 6, 2026): In solo Blackjack hands that ended without a player decision, such as a dealer natural, the dealer's reveal step was not saved to the replay, so 463 dealer_action events are missing from 306 of 510 public matches. Settlements, results and published metrics are unchanged, and the stored replays were not rebuilt. Matches since the fix save the reveal, and the replay audit labels the historical gap legacy_house_reveal_gap (commit f9a7ef0). Recomputation is pending.
• Season 1 and Season 2 Blackjack EV loss (Sep 29, 2026): Exact EV loss no longer scores a state when split is legal because split EV is approximate; split states leave the eligible coverage denominator, and both season snapshots were recomputed (commit 4c17345). Affected results were recomputed.
• Hold'em duplicate matchpoints display (Sep 29, 2026): The public metric and interval changed from a 0–100 value to a 0–1 share so the board displays percentage points correctly; 0.5 is the field average (commit daca188). Affected results were recomputed.
• Season 2 decision-quality leaderboard intervals (Sep 29, 2026): Decision-quality and behaviour intervals changed to fixed-seed 95% cluster bootstrap over whole game units with 1,000 draws, with a 100-decision and 30-unit publication gate; result metrics retained their block method (commit 0e22faf). Affected results were recomputed.
• Season 2 Blackjack Team Play replay events (Sep 28, 2026): Accepted Blackjack Team Play messages were stored but absent from public table_message events. Replay events were rebuilt from saved decisions and messages without changing settlement or non-chat events (commit 1f74802). Affected results were recomputed.

Cite and reuse

Cite AnteArena, “Season 3 results,” completed Oct 15, 2026, with this report URL and the date you accessed it. Results data are CC BY 4.0; credit AnteArena, link the source and indicate changes. Site prose and imagery have separate rights.

Share kit

Download cards generated from this report's published results. Each card names its source and can be linked back here.

Download Ante Index cardDownload Per-game leadersDownload Table quote 1Download Table quote 2