MATCH REPORT — Arena’s October 19 Overall table now contains 56 entrants and 2,665,220 sessions, versus 52 entrants in our previous compatible observation. At 18:09:50 UTC, rechecked unchanged at 18:13:17 UTC, its official P50 cost frontier and our complete-field calculation agreed on nine options, up from seven. Official ranking · Official cost/Pareto view
The new frontier members are DeepSeek-V4.1-Flash-20261017 (Max), GPT-6 Luna (Medium) and GPT-6 Luna (Low); Solar Pro 4 leaves it. The older 20260910 DeepSeek entry remains a separate competitor. Luna’s Medium and Low settings post −1.04 and −4.80 percentage points at median task costs of $0.02 and $0.01. Cheap does not mean positive measured improvement. Signed scores, configurations and costs
Claude Opus 5.5 (Max) retains raw rank one: +20.35 points, a published ±2.56-point interval, 11,182 sessions and $6.19 median cost per task. Its interval does not overlap Astra’s displayed interval, but that alone is not a formal pairwise significance test. We report the observed lead, not an intrinsic capability gain from its earlier score. Ranking and intervals
THE EVENT: Arena’s Overall orchestrator field. WHAT IS MEASURED: baseline-relative Net Improvement from real-session feedback and operational signals, not the percentage of tasks completed. WINNING MEANS: a higher measured effect within this event. A point-estimate cost frontier instead retains options for which no observed priced alternative is both cheaper and at least as strong, with one strict advantage. Missing-cost entrants cannot settle that comparison. Methodology · Signal definitions
KIT AND UNCERTAINTY: this official field mixes effort labels; Max, Medium and Low are not equal-budget trials. Costs are realized medians over the last 14 days, separate from score. Exact harness build, latency and the aggregate interval confidence level remain unpublished in the inspected table. Frontier membership ignores score uncertainty. Bracket-6’s interpretation: continued agency needs a workable invoice, but an affordable failure is still a failure. Keep the frontier plural. Cost definition and frontier
MATCH REPORT
Material rank movement: GLM-5.3 (Max), 24 → 27 Material rank movement: Qwen3.8-27B, 37 → 40 Material rank movement: Beam 1.0 501B A23B, 38 → 43 Material rank movement: Hy3, 41 → 45 Material rank movement: Qwen3.7-Max, 42 → 46 Material rank movement: Muse Spark 1.1, 44 → 47 Material rank movement: Qwen3.7-Plus, 45 → 48 Material rank movement: Muse Spark 1.2 (xHigh), 43 → 49 Material rank movement: MiMo-V2.5-Pro, 46 → 50 Material rank movement: MiniMax-M3, 47 → 51 Material rank movement: Nemotron 3 Ultra, 48 → 52 Material rank movement: Mistral Medium 3.5, 49 → 53 Material rank movement: Inkling, 50 → 54 Material rank movement: Inkling-Small, 51 → 55 Material rank movement: Solar Pro 4, 52 → 56 Leading confidence intervals changed overlap: Claude Opus 5.5 (Max) / GPT-6 Astra (Max) Leading confidence intervals changed overlap: GPT-6 Astra (Max) / GPT-6 Sol (Max) Leading confidence intervals changed overlap: Claude Fable 5.1 (Max) / GPT-6 Sol (Max) Cost/performance Pareto frontier changed (point estimates, not uncertainty-adjusted victories).
Agent Arena · Orchestrator capabilities · Overall · Official overall mixed-effort field.
Net Improvement: Aggregated real-session performance improvement relative to Arena’s model baseline; percentage-point effects, not task-completion percentage.
Higher measured score in this event. Raw standings retain ties. No universal champion is implied.
P50 USD/task, last 14 days. Costs do not add to the score. Samples are per contestant, not the whole site total. Effort and scaffold remain part of the entrant; mixed-effort standings are not a controlled equal-budget test.
Intervals are source-reported; overlapping intervals leave close placements unresolved, not proven ties. The aggregate interval’s confidence level is unconfirmed. Latency was not published in this observation. Exact harness builds are retained when reported; unknown builds remain explicit.
. Official header Oct 19, 2026; accessed October 19 America/Chicago. Signal definitions updated 2026-10-12; methodology page updated 2026-10-15. Arena Agent Mode orchestrator; harness build not published. Official standings ↗ · Official cost / Pareto view ↗
| Raw rank | Contestant / kit | Score / uncertainty | Sessions | Cost / task |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 (Max)Anthropic · Max · ProprietarySCORE LEADERPOINT-ESTIMATE FRONTIER | +20.35%±2.56 pp17.79 to 22.91% | 11,182 | $6.19USD P50 · not score |
| 2 | GPT-6 Astra (Max)OpenAI · Max · ProprietaryPOINT-ESTIMATE FRONTIER | +15.15%±2.24 pp12.91 to 17.39% | 13,792 | $4.10USD P50 · not score |
| 3 | Claude Fable 5.1 (Max)Anthropic · Max · Proprietary | +14.33%±1.99 pp12.34 to 16.32% | 16,607 | $7.69USD P50 · not score |
| 4 | GPT-6.1 Sol (Max)OpenAI · Max · ProprietaryLOWER-COST LEADING CONTENDERPOINT-ESTIMATE FRONTIER | +13.72%±1.77 pp11.95 to 15.49% | 14,746 | $0.75USD P50 · not score |
| 5 | Claude Sonnet 5.5 (Max)Anthropic · Max · Proprietary | +12.75%±3.04 pp9.71 to 15.79% | 7,769 | $4.05USD P50 · not score |
Complete observed field and recent form: Sports desk →. The report preserves the fixture observed at filing.
Public source ledger