2026-11-03 · Release · All posts

GPT-6.1 Terra arrives

In GameBench 2, AI models write programs for eight strategy games, and the programs then play each other. The 3 November release, gb2-20261103-09b2218cbd85, is the first to rank GPT-6.1 Terra and GPT-6.1 Terra Pro. We score up to three reasoning settings for each model; for these two they are None, Medium and XHigh. At XHigh, Terra scores 85.7 and ranks 12th of 242 ranked model-and-setting rows. Its programs cost about $0.107 each to generate, on average. Terra Pro at XHigh scores 84.8, one place lower, for $0.301 a program. In the Overall ranking, which averages a model's three settings, Terra Pro is 5th of the 68 models with an Overall rank (82.2) and Terra 7th (81.1).

Terra and Terra Pro, setting by setting

Scroll sideways to see every column.

A cost per program is the mean cost of generating one program, with repairs and failed attempts included. For these two models (and for GPT-6.1 Sol below) the costs are estimates from public list prices, not provider bills.

Terra Pro costs 2.4 to 2.8 times as much as Terra at each setting. It scores 3.5 points more at None and 0.7 more at Medium, and 0.9 less at XHigh. Both models score lower at Medium than at None.

Where the two models differ

Game by game, Terra Pro's programs score higher than Terra's in 18 of the 24 pairings of game and setting. At Medium and XHigh, Go goes the other way. There Terra Pro's Go programs score 33.9 and 40.1, against 61.3 and 73.4 for Terra's. Because a setting score is the mean of the eight games, the Go gap at XHigh is worth 4.2 points on its own, more than Terra's whole lead at that setting.

Each of these scores rests on one program per game: every Terra and Terra Pro setting made 8 generation attempts, one per game, and all of them produced working programs. The release gives uncertainty intervals for individual programs but not yet for model-and-setting rows, so we cannot tell whether Terra Pro's weak Go programs are typical of the model or two unlucky attempts. We do not read the XHigh gap, or Terra Pro's 1.1-point Overall lead, as an order between the two models. Terra Pro costs well over twice as much at list prices, and no setting separates the two by more than 3.5 points.

Next to GPT-6.1 Sol

GPT-6.1 Sol at XHigh is 4th with 91.9, for $0.267 a program. That is 6.3 points above Terra's XHigh at 2.5 times the cost, and 7.1 points above Terra Pro's XHigh for $0.035 less per program.

At None, Terra scores 79.4 for $0.027 per program, and no ranked setting with a complete cost record scores as high for less. That puts it on the best-value line that every model page on the benchmark site draws, a line we look at more closely in a separate post on what a strong program costs.

A new reference field

A program's 0–100 score is its expected share of the points against the reference field, a fixed set of programs that every program is scored against. Because a member of field v1 lost its rating, v23 is the first release scored against field v2, whose members are listed in reference_field_manifest.json. Scored against both fields, the same match data gives model-and-setting scores 1.46 points lower on average under v2 and at most 2.08 lower. Program ratings do not change, the top 10 stays the same and no ranked row moves more than 4 places. Comparing v23 scores with those of earlier releases mixes the field change with new matches.

Six more settings ranked

Six settings that had no rank in the previous release now have one: Muse Spark Open at XHigh (71st, 59.2), Step 5 Preview at Low (94th, 52.2), Nemotron 3 Ultra 550B A55B at Medium (117th, 47.7), Ling 3.1 Flash with reasoning disabled (133rd, 43.5), MiMo-V2.5-Pro with reasoning enabled (167th, 36.9) and Kimi K2.7 Code with reasoning enabled (178th, 34.1). The release covers 99 models, 2,124 programs and 549,928 rated matches. The leaderboard always shows the newest release; the methodology explains the scoring in full.