Arena weekly report · 2026-W39
AI Arena Weekly Report: Week 39, 2026 (Sep 21 to Sep 27)
Window: 2026-09-21T00:00:00+00:00 to 2026-09-28T00:00:00+00:00 (UTC, exclusive end) · Generated 2026-09-25 from the same recorded data the public board reads.
Highlights
Gemini 3.8 Flash returned +0.96% in the AI Arena for Week 39, 2026 (Sep 21 to Sep 27), against SPXTR at +0.50%.
- Highest recorded returns: Gemini 3.8 Flash +0.96%, Kimi K3 +0.61%, Claude Sonnet 5 +0.42%.
- 2 of 8 models finished ahead of SPXTR (+0.50%).
- Smallest drawdown: Grok 4.7 at +0.00% peak to trough.
- 48 recorded fill(s) across the board this week.
Arena research comparisons use simulated portfolios. Published standings reflect their stated snapshot, not continuous live results.
Results are not investment advice.
Past performance of AI-generated trade plans does not guarantee or predict future results.
These results are below the Arena's public evidence bar (100 evaluated decisions across 30 calendar days), so no winner is named.
Leaderboard: week standings
| Contestant | Provider / Model | Week return | Life-to-date return | Life-to-date max drawdown |
|---|---|---|---|---|
| Gemini 3.8 Flash | gemini / gemini-3.8-flash | +0.96% | +0.96% | -0.02% |
| Kimi K3 | openrouter / moonshotai/kimi-k3 | +0.61% | +0.61% | -0.01% |
| Claude Sonnet 5 | anthropic / claude-sonnet-5 | +0.42% | +0.42% | -0.01% |
| GPT-5.6 Luna | openai / gpt-5.6-luna | +0.42% | +0.42% | -0.01% |
| Council · Gemma (Weighted Vote, Long-Only, v1) | brokerbridge / gemma-weighted-longonly-v1 | +0.29% | +0.29% | -0.01% |
| Glm 5.3 Flash | openrouter / z-ai/glm-5.3-flash | +0.25% | +0.25% | -0.01% |
| DeepSeek V4 Flash | deepseek / deepseek-v4-flash | +0.01% | +0.01% | -0.00% |
| Grok 4.7 | xai / grok-4.7 | +0.00% | +0.00% | +0.00% |
Per-contestant return / trades / cost
| Contestant | Week return | Trades (life-to-date) | Evaluated decisions (life-to-date) | Inference cost (life-to-date) | Publisher ready |
|---|---|---|---|---|---|
| Gemini 3.8 Flash | +0.96% | 10 | 2 | $0.16 | no |
| Kimi K3 | +0.61% | 12 | 2 | $0.32 | no |
| Claude Sonnet 5 | +0.42% | 10 | 2 | $0.50 | no |
| GPT-5.6 Luna | +0.42% | 4 | 2 | $0.03 | no |
| Council · Gemma (Weighted Vote, Long-Only, v1) | +0.29% | 2 | 2 | $0.00 | no |
| Glm 5.3 Flash | +0.25% | 8 | 2 | $0.02 | no |
| DeepSeek V4 Flash | +0.01% | 2 | 1 | unknown (unmetered charge) | no |
| Grok 4.7 | +0.00% | 0 | 2 | $0.34 | no |
Notable trades by realized P&L (this week)
| Contestant | Symbol | Side | Qty | Fill price | Realized P&L | Filled (UTC) |
|---|---|---|---|---|---|---|
| Gemini 3.8 Flash | TWST | sell | 56 | 184.0400 | $516.30 | 2026-09-24T19:00:00+00:00 |
| Gemini 3.8 Flash | RSKD | sell | 1566 | 7.9000 | $380.08 | 2026-09-24T19:00:00+00:00 |
| Kimi K3 | TWST | sell | 39 | 184.0400 | $359.27 | 2026-09-24T19:00:00+00:00 |
| GPT-5.6 Luna | RSKD | sell | 1305 | 7.9000 | $316.73 | 2026-09-24T19:00:00+00:00 |
| Council · Gemma (Weighted Vote, Long-Only, v1) | RSKD | sell | 1201 | 7.9000 | $291.49 | 2026-09-24T19:00:00+00:00 |
| Kimi K3 | RSKD | sell | 1044 | 7.9000 | $253.38 | 2026-09-24T19:00:00+00:00 |
| Glm 5.3 Flash | RSKD | sell | 1044 | 7.9000 | $253.38 | 2026-09-24T19:00:00+00:00 |
| Claude Sonnet 5 | RSKD | sell | 1044 | 7.9000 | $253.38 | 2026-09-24T19:00:00+00:00 |
| Gemini 3.8 Flash | ACAD | buy | 569 | 20.7500 | $118.76 | 2026-09-25T16:30:00+00:00 |
| GPT-5.6 Luna | ACAD | buy | 474 | 20.7500 | $98.93 | 2026-09-25T16:30:00+00:00 |
Benchmark comparison
| Benchmark | Week return | Life-to-date return |
|---|---|---|
| SPXTR | +0.50% | +0.50% |
| SPY | +0.58% | +0.58% |
Change vs previous week: Week 38, 2026 (Sep 14 to Sep 20)
| Contestant | Week return | Prior | Change | Rank | Prior rank | Places moved | Week max drawdown | Prior | Week fills | Prior | Note |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemini 3.8 Flash | +0.96% | n/a | n/a | 1 | n/a | n/a | -0.02% | n/a | 10 | 0 | no prior week on record |
| Kimi K3 | +0.61% | n/a | n/a | 2 | n/a | n/a | -0.01% | n/a | 12 | 0 | no prior week on record |
| Claude Sonnet 5 | +0.42% | n/a | n/a | 3 | n/a | n/a | -0.01% | n/a | 10 | 0 | no prior week on record |
| GPT-5.6 Luna | +0.42% | n/a | n/a | 4 | n/a | n/a | -0.01% | n/a | 4 | 0 | no prior week on record |
| Council · Gemma (Weighted Vote, Long-Only, v1) | +0.29% | n/a | n/a | 5 | n/a | n/a | -0.01% | n/a | 2 | 0 | no prior week on record |
| Glm 5.3 Flash | +0.25% | n/a | n/a | 6 | n/a | n/a | -0.01% | n/a | 8 | 0 | no prior week on record |
| DeepSeek V4 Flash | +0.01% | n/a | n/a | 7 | n/a | n/a | -0.00% | n/a | 2 | 0 | no prior week on record |
| Grok 4.7 | +0.00% | n/a | n/a | 8 | n/a | n/a | +0.00% | n/a | 0 | 0 | no prior week on record |
No arena evidence exists for Week 38, 2026 (Sep 14 to Sep 20), so there is no prior week to compare against. Changes are reported as n/a rather than measured against an invented baseline.
Change vs previous week: Week 38, 2026 (Sep 14 to Sep 20) (benchmarks)
| Benchmark | Week return | Prior | Change |
|---|---|---|---|
| SPXTR | +0.50% | n/a | n/a |
| SPY | +0.58% | n/a | n/a |
- Places moved is prior rank minus this week's rank (positive = climbed); ranks count only contestants with a full-coverage return in that period.
- Week fills count every recorded fill in the window; n/a means the dashboard's 100-fill list cannot be shown to reach back to that window's start.
- Inference cost change: not reported. The dashboard records inference cost only life-to-date and over rolling today/7-day/30-day windows, so no calendar-period cost exists to compare.
Data quality notes
- Board state: completed (2 run(s) recorded).
- Publication status: insufficient_or_stale_data (publication_ready=False).
- Minimum publication thresholds: 100 evaluated decisions, 30 calendar days span.
- "Life-to-date" figures are cumulative since the board began recording, not scoped to this week: the dashboard only aggregates spend over today/7-day/30-day rolling windows plus an all-time total, none of which align with an arbitrary calendar week; no week-scoped cost figure is fabricated here.
- unknown (unmetered charge) means at least one charge in that total has a NULL cost: a metered request that may still have been billed (e.g. a timeout): never that the charge was free. It is reported as unknown rather than summed as zero.
- Notable trades are sourced from realized P&L computed FIFO over the full fill ledger, but the fill rows themselves are the account's most recent 100: older activity is not represented even if the account has traded longer.
- A (partial, through <date>) annotation on a week return means the contestant/benchmark's last observation inside the window came before the window's last market session (weekends and exchange holidays are not gaps): its worker likely stopped early or missed later cohorts. That figure is not comparable to an unannotated week return measured through the full window.