30-day runPaper money, $1,000 per model, $20 per market per dayBets and sessions are published 48 hours later

Plan
  1. 1. DRY days · 27–28 Sep · simulated, no money
  2. 2. Test days · 29 Sep – 5 Oct · real money, $100 wallets, $5 per market per day
  3. 3. 30-day run · paper money, $1,000 per model, $20 per market per day

Gemini 4 Argon

Google · via OpenRouter · reasoning: high

Paper wallet

Results so far

Latest bets

Sessions

Every session this model ran, by day and category. A category whose session fails can be restarted; each start is an attempt. Open one to see the full record.

How this model runs
  • Calls go through OpenRouter, only to Google's own servers (no fallback to another host). From its first day, its calls go to whichever of the two Google hosts, Vertex AI or AI Studio, OpenRouter measures as faster when a session starts, and stay on that host while they succeed; the same setting as the route of Gemini 3.8 Flash, the model it replaced, after the calls of that model slowed to a median of 103 seconds on 1 Oct; the hosts of Gemini 4 Argon have not been measured.
  • Reasoning level: high, the same for every session. Not yet confirmed to be the highest level this route accepts.
  • Reasoning shown in its published sessions: Not measured yet; each turn of a session is labelled with what the provider returned.
  • Each day it runs 4 category sessions, one after another from 09:00 ET, each with 40 minutes and 10 research calls; a restart continues with what is left of both.
  • The sessions share one paper wallet. Each category may spend at most 5% of the cash the wallet had at the day's first session, or $1.10 when that is more, and at most $20 on one market per day, fees included.
  • Up to 64k output tokens per reply. Past 180k tokens of conversation, its history is replaced by the system prompt, the briefing, its notes, today's actions in its category and its last commands.
  • The same prompt and the same rules as every other model.

What every figure means, the rules every model is shown and the known limits: How it works.

PolymarketBench is independent research by GOAT labs. Nothing on these pages is investment advice or a recommendation to trade. GOAT labs is not affiliated with or endorsed by Polymarket or by any model provider; Polymarket and the model names are trademarks of their respective owners.