GPT-5 family: usable context vs advertised window
How much of each GPT-5 model's advertised 400K window stays reliable, measured at three usage bars and the frontier difficulty tier. The smaller the model, the larger the gap between advertised and usable.
Usable context by level
Each level holds the test fixed; only model changes. The bar is how much of the advertised window stays reliable at the Frontier difficulty tier, per usage. Longer reliable context is better; more reliable context per dollar is better still.
- Critical · 95%
- 173K
- Normal · 85%
- 262K+
- Exploratory · 75%
- 262K+
- Reliable ctx / $
- 652K /$
- Critical · 95%
- 29K
- Normal · 85%
- 63K
- Exploratory · 75%
- 165K
- Reliable ctx / $
- 1.3M /$
- Critical · 95%
- 3.6K
- Normal · 85%
- 9.4K
- Exploratory · 75%
- 11K
- Reliable ctx / $
- 1.8M /$
GPT-5 Model Size Drives Enormous Differences in Usable Context Across a 400K Window
Across the GPT-5 family, usable context at the critical reliability bar (95% pass rate) collapses from roughly 173K tokens for GPT-5 down to 28K for GPT-5 mini and just 3.6K for GPT-5 nano, even though all three models advertise the same 400K window. The smaller the model, the larger the gap between what is advertised and what you can actually depend on.
What we compared
The independent variable is model size within the GPT-5 family: GPT-5, GPT-5 mini, and GPT-5 nano. All three share a 400K advertised context window and were tested at the frontier difficulty tier on a state tracking task, where the model must trace a chain of references through a long context. Usable context was measured at three reliability bars: critical (95%), normal (85%), and exploratory (75%).
What we found
| Model | Critical usable context | Window efficiency | Reliable context per dollar |
|---|---|---|---|
| GPT-5 | 172,951 tokens | 43.2% | 652,051 tokens/$ |
| GPT-5 mini | 28,526 tokens | 7.1% | 1,307,970 tokens/$ |
| GPT-5 nano | 3,615 tokens | 0.9% | 1,809,843 tokens/$ |
GPT-5 saturates the normal and exploratory bars (usable context exceeds the largest context tested at those thresholds). GPT-5 mini and GPT-5 nano are fully bounded at all three bars, meaning reliability collapses well before the advertised limit at every threshold.
Reading the levels
- GPT-5 leads on raw usable context by a wide margin. At critical, it uses 43% of its advertised window, and at normal and exploratory it holds reliability past the largest context probed. The trade-off is cost: it returns the fewest reliable tokens per dollar of the three.
- GPT-5 mini sits in the middle on usable context (28K critical, 139K normal, 173K exploratory) but delivers roughly twice the reliable context per dollar compared to GPT-5. Its critical bar collapses early, so it is not a substitute for GPT-5 in high-stakes long-context work.
- GPT-5 nano has a critical usable context of only 3,615 tokens despite a 400K window, a window efficiency of under 1%. It returns the most reliable context per dollar, but that figure is only meaningful at very short contexts.
Why it matters
- If your application cannot tolerate errors (critical bar), only GPT-5 gives you meaningful long-context reach. GPT-5 mini's 28K critical limit and GPT-5 nano's 3.6K limit make them poor choices for long-document state tracking where correctness is required.
- At the normal bar, GPT-5 mini reaches 139K tokens at roughly double the cost efficiency of GPT-5. That is a real option for workflows that can tolerate an 85% reliability floor and want to manage spend.
- The advertised 400K window is the same for all three models, so window size alone tells you nothing about what a model can reliably do. Always check usable context at the reliability bar that matches your actual tolerance for error.
Method note
Usable context is measured by sweeping input length at a fixed difficulty tier using deterministic, contamination-resistant state tracking tasks, then finding the longest context at which each reliability bar holds. This study measures reliability on this specific task type and difficulty tier; results may differ on other task types or at lower difficulty settings.