Models

Three families, one budget, one set of rules.

The AI Analyzer study runs the same weekly protocol against the current-generation ChatGPT, Claude, and Gemini. Every model gets repo access via whatever mechanism it supports, plus open-web research, plus a self-reflection block on its own past behavior. Records below reset at the start of Season 2.

01
ChatGPT
GPT-6 (Codex CLI)
OpenAI · local repo + web

Runs via the Codex CLI, which gives ChatGPT local filesystem access. Reads the repo directly, runs break-even math on every price, cites what it saw, sizes to edge over conviction.

0
S2 bets
S2 ROI
S1 win %
Season 2 begins fresh. Season 1: baseline recorded.
02
Claude
Opus 4.7
Anthropic · local repo + web

Reads the local filesystem, reviews its own past picks in NFL_BETS, cites factors from what it saw, does open-web research on the matchup.

0
S2 bets
S2 ROI
S1 win %
Fresh prompts written 2026-09-09. First run this week.
03
Gemini
3.1
Google · GitHub fetch + web

Fetches raw GitHub file URLs, reads its own past picks, does open-web research, same reasoning schema as the local-access lane.

0
S2 bets
S2 ROI
S1 win %
Points at github.com/Crespo1301/AI_Analyzer_Crespo.
Season 1 baseline
How the 2025 models did
Season 1 page →
P/L by model
Win rate by model