GPT-5
At the Frontier difficulty, GPT-5 stays reliable to 173K of its 400K advertised window for critical use (43% of the window), at 652K /$ of reliable context.
Frontier difficulty
Hardest tasks, tuned to bend the strongest models
- Critical ≥95%
- 173K *
- CI 2.9K to 262K
- Right 95% of the time. Use it where a wrong answer is costly.
- Normal ≥85%
- 262K+
- Right 85% of the time. A person reviews the output.
- Exploratory ≥75%
- 262K+
- Right 75% of the time. Fine for drafts and ideation.
Standard difficulty
A moderate task difficulty
- Critical ≥95%
- 262K+ *
- Right 95% of the time. Use it where a wrong answer is costly.
- Normal ≥85%
- 262K+ *
- Right 85% of the time. A person reviews the output.
- Exploratory ≥75%
- 262K+
- Right 75% of the time. Fine for drafts and ideation.
Light difficulty
Lighter tasks, where cheaper and self-hosted models still bend in window
- Critical ≥95%
- 262K+ *
- Right 95% of the time. Use it where a wrong answer is costly.
- Normal ≥85%
- 262K+ *
- Right 85% of the time. A person reviews the output.
- Exploratory ≥75%
- 262K+
- Right 75% of the time. Fine for drafts and ideation.
Usable context is the longest context a model stays at or above a usage bar, read from a fit that borrows strength across the difficulty tiers so the higher bars resolve without deep per-cell sampling. A star marks a bar whose lower confidence bound is not resolved at this depth. Value is reliable context per dollar of input price; self-hosted models are priced by amortized GPU-hours.
Usable context above is the primary measure. As a secondary, difficulty-axis view (how hard a coupled reasoning task it sustains as context grows), GPT-5 scores 73.2/100.