Near the top, generation cost varies by 6.4×
GPT-5.6 Sol at XHigh is almost tied for first in the current GameBench 2 standings. It scores 72.4, only 1.3 points behind Claude Fable 5 at XHigh. But their estimated mean generation costs are $10.06 and $1.58 per game-specific player. Sol costs 6.4× as much for a result that is very close at the top of the table.
Sol is not an isolated case. Across GPT-5.6 Sol, Terra, and Luna, XHigh produces the best score—and costs about 22×, 49×, and 88× as much as Medium. That family-wide pattern points to a more interesting question than the cross-provider price gap alone: is GPT-5.6's XHigh setting spending far more reasoning than this task needs?
TL;DR.
1. GPT-5.6 Sol XHigh trails the leader by only 1.3 points but costs 6.4× as much to generate at standard list prices.
2. Within GPT-5.6, XHigh improves the score by 14.0 to 25.7 points over Medium, but raises estimated generation cost by about 22× to 88×.
3. The output-token evidence is consistent with XHigh not yet being well calibrated for this task. That is a plausible explanation, not a proven cause.
The top six do not form a price ladder
| Rank | Model | Reasoning | Score | Estimated mean cost per game |
|---|---|---|---|---|
| 1 | Claude Fable 5 | XHigh | 73.7 | $1.58 |
| 2 | GPT-5.6 Sol | XHigh | 72.4 | $10.06 |
| 3 | Claude Fable 5 | Medium | 69.8 | $0.50 |
| 4 | Claude Opus 4.8 | XHigh | 68.6 | $0.93 |
| 5 | GPT-5.5 | XHigh | 67.6 | $2.74 |
| 6 | GPT-5.4 | XHigh | 66.8 | $2.64 |
The GPT-5.6 family repeats the pattern
| Model | Medium score | XHigh score | Score gain | Medium cost | XHigh cost | Cost ratio |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 52.3 | 72.4 | +20.1 | $0.47 | $10.06 | 21.5× |
| GPT-5.6 Terra | 37.9 | 51.9 | +14.0 | $0.17 | $8.17 | 48.8× |
| GPT-5.6 Luna | 29.3 | 55.0 | +25.7 | $0.04 | $3.16 | 87.5× |
XHigh clearly improves playing strength for all three models. The gains are meaningful: 20.1 points for Sol, 14.0 for Terra, and 25.7 for Luna. But the cost increase is much larger than the score increase. For these GPT-5.6 results, moving from Medium to XHigh is not a small premium for a small refinement; it is a different spending regime.
The token counts point to calibration
The cost jump follows an unusually large jump in recorded output-token use. Sol averages about 13,800 output tokens at Medium and 334,000 at XHigh. Terra rises from about 9,400 to 542,000, while Luna rises from about 4,600 to 532,000. Those are roughly 24×, 58×, and 117× increases. Because output tokens are a major part of generation price, this helps explain why the XHigh rows cost so much more.
This pattern is consistent with GPT-5.6's XHigh setting not yet being well calibrated for this code-generation task: it may be allocating far more reasoning than is useful at the margin. The benchmark cannot prove that diagnosis. The source could be model behavior, the provider's effort policy, request routing, or an interaction with this particular task. Future results will show whether the token use falls as the models or their XHigh behavior mature.
Claude Fable 5 shows a different tradeoff
Claude Fable 5 moves from 69.8 at Medium to 73.7 at XHigh while its estimated mean cost rises from $0.50 to $1.58, about 3.1×. Its Medium result also includes one failed player: the seven working players average 72.2, but the missing eighth lowers the headline score to 69.8. Even with that reliability difference, Fable's cost curve is far less steep than the GPT-5.6 curves in this snapshot.
Cross-provider comparisons need care: XHigh is each model's highest supported setting, not a standardized amount of computation. Still, the within-family Medium-to-XHigh comparisons make the GPT-5.6 pattern hard to dismiss as list-price differences alone.
What these results show
GPT-5.6 XHigh buys substantial playing strength, and Sol reaches almost the very top. It currently does so at a striking generation premium. The best reading is not that XHigh is wasteful, but that its extra strength is expensive and its token allocation may still be immature for this task. Medium is dramatically cheaper; XHigh is the choice when the additional score matters more than generation cost.
Costs here are estimated averages for generating one game-specific player, including repeat or repair attempts. They use dated public list prices, do not affect rank, and are not provider invoices or deployment costs. See the current leaderboard for updated results, or Introducing GameBench 2 for the benchmark overview.