Reverse-engineering Codex credit limits
Starting from the Codex pricing page and rate card: GPT-5.4 costs 62.5 credits per 1M input tokens vs $2.50 on the API. 2.50/62.5 = $0.04 per credit. This holds for cached (0.25/6.25), output (15/375), and other models.
The usage tables show Plus has 100 max GPT-5.4 messages per 5h, 350 max GPT-5.4-mini, 150 max GPT-5.3-Codex. The legacy rate card gives approximate credit costs: ~7, ~2, ~5 respectively. 100×7 = 350×2 ≅ 150×5 = 700-750 credits. Using 750 as the baseline: $30/5h.
Pro 5× upper limits are exactly 5× Plus (500 vs 100, 1750 vs 350, 750 vs 150), so 3,750 credits/5h. Pro 20× is 15,000. The current promos make $100 effectively 10× (7,500) and $200 effectively 25× (18,750) until May 31 2026.
Weekly limits aren't published ("additional weekly limits may apply"). Assuming 6 maxed sessions hit the weekly cap, and 4.33 weeks/month:
| Plan | Credits/5h | Credits/month | API equiv/month |
|---|---|---|---|
| Plus ($20) | 750 | 19,485 | $779.40 |
| Pro $100 (5×) | 3,750 | 97,425 | $3,897 |
| Pro $200 (20×) | 15,000 | 389,700 | $15,588 |