GPT-5-nano: does more reasoning buy more usable context?
GPT-5-nano measured at low, medium, and high reasoning effort, at a fixed difficulty tier. More reasoning shifts both usable context and the cost of a reliable result, so the cheapest level that clears your length can beat the deepest one.
Usable context by level
Each level holds the test fixed; only reasoning effort 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%
- 4.5K
- Normal · 85%
- 9.0K
- Exploratory · 75%
- 13K
- Reliable ctx / $
- 806K /$
- Critical · 95%
- 3.6K
- Normal · 85%
- 9.4K
- Exploratory · 75%
- 11K
- Reliable ctx / $
- 1.8M /$
- Critical · 95%
- 3.3K
- Normal · 85%
- 4.4K
- Exploratory · 75%
- 5.8K
- Reliable ctx / $
- 4.3M /$
More Reasoning Effort Buys GPT-5-nano More Usable Context, but at a Steep Cost-per-Token Price
At the frontier difficulty tier, raising GPT-5-nano from low to high reasoning effort roughly doubles the usable context at the critical bar (3,327 tokens to 4,492 tokens) and more than doubles it at the exploratory bar (5,793 tokens to 13,004 tokens). That sounds like a clear win, but the model's advertised window is 400,000 tokens, and even at high effort the critical bar clears only 1.1% of that window. The cost story runs in the opposite direction: low effort delivers 4.27 million reliable tokens per dollar versus 806,000 at high effort, a 5x gap in value.
What we compared
The independent variable is reasoning effort, tested at three levels: low, medium, and high. All three runs use GPT-5-nano on the same hosted deployment, the same frontier difficulty tier, and the same three reliability bars: critical (95% pass rate), normal (85%), and exploratory (75%). The advertised context window is 400,000 tokens in every case, so the only thing changing is how much reasoning the model applies.
What we found
| Effort | Critical (tokens) | Window efficiency | Reliable context per dollar |
|---|---|---|---|
| High | 4,492 | 1.12% | 806,164 |
| Medium | 3,615 | 0.90% | 1,809,843 |
| Low | 3,327 | 0.83% | 4,271,643 |
All bars are bounded at every effort level; no level saturates the tier and no level floors.
Reading the levels
- High effort leads on usable context at every bar, most noticeably at exploratory (13,004 tokens vs. 5,793 at low). The gain is real but modest at the critical bar, only about 1,165 tokens over low effort.
- Medium effort sits in an awkward middle: it beats low on normal and exploratory context but trails high on all three bars, and it costs more than low without delivering high's ceiling.
- Low effort is the clear winner on value. It returns 4.27 million reliable tokens per dollar, more than 5x high effort, making it the rational choice whenever your context fits inside its tighter ceiling.
Why it matters
- If your workload requires state tracking across more than roughly 9,000 tokens, only high effort clears the normal bar at that length; low effort fails well before that point (4,390 tokens at normal).
- If your context fits inside 3,300 to 4,400 tokens, low effort is sufficient at every bar and costs a fraction of high effort. Paying for more reasoning buys nothing in that range.
- The 400,000-token advertised window is effectively decorative here. Even the best configuration (high effort, exploratory bar) uses 3.3% of it. Do not plan workloads around the advertised window for this task type.
Method note
Usable context is measured by sweeping input length at fixed difficulty with deterministic, contamination-resistant state-tracking tasks, then finding the longest context that holds each reliability threshold. This study isolates reasoning effort on one model at one difficulty tier; it does not generalize to other task types or other models.