Reports

AI Game Dev Benchmarks, November 2026: 3D, Creative Writing and World Building

Everything we published in our October benchmark baseline happened before five major model drops and a full rescaling of Artificial Analysis's core suites [1][2]. This update pulls fresh 12 November 2026 numbers across 3D scene generation, creative writing, lore grounding, spatial world building, coding agents, and tactical game play.

By the CreateGame.ai editorial teamNovember 12, 202619 min read

What changed since 9 October

Our October game-development report snapped Artificial Analysis when Intelligence Index v4.3.2 placed Claude Opus 5.5 and GPT-6 Astra at the top. Between 9 October and 12 November 2026, two things happened at once: six significant model checkpoints landed, and Artificial Analysis rolled over to Intelligence Index v5.0.1 alongside Coding Agents Index (FrontierCode v1.1), AA-Omniscience v1.1, an updated MLCR-AA multi-turn judge panel (6 Nov), and the new AA-3D benchmark [1][2][3][9].

Do not compare October's v4.3.2 numbers directly against November's v5.0.1 column. To keep comparisons clean, we pulled every model under the current v5.0.1 rubric on 12 November 2026 and cross-checked general capabilities against Epoch AI's latest Capabilities Index (ECI) [2][12].

Three pricing and speed shifts matter for studio budgets right now:

• Flagship accuracy still costs flagship money, especially at max reasoning. Claude Fable 5.2 (max) takes #1 on both Intelligence Index v5.0.1 (64.56) and ECI (173.44), but a full benchmark run burns $26.54 at list $8/$40 per million input/output tokens (97.7 tok/s) [2][12]. Dialing down to Claude Fable 5.2 (high) retains 61.82 on v5.0.1 while cutting full-suite run cost nearly threefold to $9.08.
• Two mid-tier checkpoints broke the $1–$3 ceiling. GPT-6.1 Terra (max) (released 26 Oct, $1/$5 list) scores 51.33 on v5.0.1 at just $0.96 per run, and Gemini 4 Hydrogen (released 12 Nov, $1.50/$7.50) hits 53.92 at xhigh (198.2 tok/s) or 50.79 at high while streaming at 235.5 tok/s [2]. Notice Epoch lists Gemini 4 Carbon (169.76 ECI) with unreleased weight metadata despite its public 28 October API pricing ($4/$20 list on AA) [2][12].
• Open weights jumped nearly 14 points in a single release. Kimi K3.1 Pro (released 23 Oct, 1M context window) reaches 46.45 on Intelligence Index v5.0.1, outscoring DeepSeek V4.1 Flash 1017 (max) (32.59) and GLM-5.3 (max) (31.65) by roughly 14 points and catching last generation's closed frontier [2]. Meta's Muse Wonder Full (max) scores 50.08 overall, though AA does not yet publish API pricing or throughput for it.

3D benchmarks for games and engines

Until recently, tracking 3D capability meant guessing from web-app Elo. Now two public leaderboards test actual 3D scene output side by side: AA-3D on Artificial Analysis and DesignArena's 3d-design and blender tracks. They reward very different strengths [3][4][5].

AA-3D tests 25 models across 55 written scene briefs, from playable pinball tables and spinning black holes to the Colosseum, using Artificial Analysis's open-source Stirrup agent in an offline Three.js runtime, ranked by blind pairwise human preference [3]. Anthropic sweeps the podium:

• Claude Fable 5.2 (max) reaches 1271.4 Elo (95% CI 1246–1297), a 68-point lead over Claude Sonnet 5.5 (max) (1203.7, CI 1176–1232) and Claude Opus 5.5 (max) (1202.4, CI 1174–1231) [3].
• Sub-category win rates inside AA-3D uncover a useful quirk: Claude Opus 5.5 (max) actually edges out Fable 5.2 (max) slightly inside the games prompt subset (84.6% balanced win rate vs. 82.7%), while Fable 5.2 runs away with stories (83.7%) and landmarks (84.3%) [3]. If your workload is mostly Three.js micro-games at half the output token price ($20/1M vs. $40/1M), Opus 5.5 remains surprisingly competitive.
• Kimi K3.1 Pro (1112.1 Elo; 67.2% games win rate, 75.4% landmarks) is the clear open-weights leader, beating MiMo-V2.6-Pro (982.9) and Qwen3.8 Max (954.0) by over 129 Elo and clearing Thinking Machines' Inkling 2 Small (1027.3) [3].

Notice the missing entrants before crowning a single winner: neither Gemini 4 Carbon nor Gemini 4 Hydrogen nor GPT-6.1 Terra had been rated on AA-3D as of 12 November [3].

DesignArena fills some of that gap with open community voting across broader browser 3D (3d-design, 180 models) and native Python Blender scene scripting (blender, 21 models) [4][5]:

Two findings stand out when comparing AA-3D against DesignArena:

1. Astra and Carbon show up much stronger on DesignArena than Gemini 4 Argon did on AA-3D. On DesignArena's 3d-design board, GPT-6 Astra (max) sits at 1525 Elo (77.3% win rate over 384 battles, 1201.7s average generation), nine points ahead of Claude Fable 5.2 (1516 Elo, 76.8% WR over 280 battles, 851.6s) [4]. Meanwhile Gemini 4 Carbon posts a 68.9% win rate (1426 Elo over 399 battles) while returning results in 240.1 seconds, roughly four to five times faster than the two leaders, and Kimi K3.1 Pro (1446 Elo, 71.0% WR) again clears Claude Sonnet 5.5 (1401 Elo) [4].
2. In Blender, native 3D mesh generators match frontier code-generation LLMs in a fraction of the wait. Claude Opus 5.5 (max) (1413 Elo, 71.7% win rate) and GPT-6 Astra (max) (1412 Elo, 68.8% win rate) edge out Gemini 4 Argon (1378 Elo, 25.6 minutes per scene) and GPT-6 Astra base (1363 Elo, 21.3 minutes) [5]. Sitting right between them at #4 is Meshy 7.1 at 1371 Elo and a 71.8% win rate in 88.3 seconds [5]. LLMs give you editable scene trees and procedural Python modifiers; dedicated generators give you silhouette and texture iteration speed when blocking out levels.

Creative writing, dialogue and RPG prose

As we noted in our AI RPG report, prose quality depends heavily on who is judging and whether verbosity gets rewarded. On 12 November, EQ-Bench Creative Writing v3 (145 models, LLM-judged pairwise Elo plus rubric sub-metrics) tells a noticeably different story from coding leaderboards [6].

GPT-6 Astra holds #1 overall at 2176.9 Elo (15.94 rubric score, 8.41 slop score), with open-weights Kimi K3.1 Pro (*kimi-k3.1-pro) only 4.3 Elo behind at 2172.6 Elo [6]. Look inside EQ-Bench's diagnostic columns before picking either blindly:

Three patterns jump off that table for narrative designers:

• Open weights caught up on style, not slop discipline. Kimi K3.1 Pro earns the highest rubric score in the top ten (16.63, ahead of GLM-5.3 at 16.44 and Fable 5.1 at 16.27), yet still triggers more cliché phrasing (9.64 slop) than Claude Opus 5 (6.59) [6]. If you generate NPC barks or item descriptions locally, plan a strict banned-phrase post-pass.
• Opus 5.5 dropped sharply vs. older Opus 5 and Fable 5.1 on prose. Claude Opus 5.5 sits down at #8 (2050.9 Elo, 10.43 slop), 84 Elo below Claude Opus 5 (2135.0 Elo, 6.59 slop) and 104 Elo below Claude Fable 5.1 (2155.3) [6]. Higher coding and reasoning scores do not guarantee cleaner dialogue.
• Most long-form creative writing boards have stalled. Neither Claude Fable 5.2 nor Gemini 4 Carbon nor Gemini 4 Hydrogen has appeared on EQ-Bench Creative Writing v3 yet [6]. Worse, when we checked Epoch's external benchmark archive on 12 November, Lech Mazur Writing still ends in August 2025 (gpt-5-2025-08-07_medium at 8.6), Fiction.liveBench (120k_tokens recall) stops at January 2026 (kimi-k2.5 at 0.781), and EQ-Bench's own Longform board has not added September–November releases past claude-sonnet-4-6 [13][14]. Creative writing remains the least consistently updated part of AI benchmarking.

Worldbuilding, lore grounding and spatial consistency

In our worldbuilding guide, we argued that lore assistants fail in three predictable ways: inventing false canon confidently, dropping details across multi-turn sessions, and breaking spatial or causal logic inside quests and maps. November gives us quantitative proxies for all three, and they expose sharp lab personality differences.

First, AA-Omniscience v1.1 (expanded to 3,000 factual questions) separates raw recall from abstention discipline [2]:

• Anthropic and OpenAI maximize recall and accept higher hallucination rates. Claude Fable 5.2 (max) leads overall AA-Omniscience (66.8) with 79.1% accuracy, yet only a 40.7% non-hallucination rate: when unsure, it prefers guessing [2]. Claude Opus 5.5 (max) (65.5% accuracy, 49.7% non-hallucination) and GPT-6 Astra (max) (61.6% accuracy, 50.2% non-hallucination) show the same willingness to guess [2]. That behavior helps ideation and risks canon corruption if left unreviewed.
• Google's Gemini 4 family and Meta's Muse Wonder heavily favor abstention. Gemini 4 Carbon (high) scores 58.3% accuracy alongside an 88.6% non-hallucination rate (rising to 95.8% at medium), while today's Gemini 4 Hydrogen (xhigh) hits 91.5% non-hallucination at 50.5% accuracy and Muse Wonder Full (max) sits at 82.4% non-hallucination [2]. For an in-editor wiki search or canon QA checker that must refuse to invent facts missing from the setting bible, Carbon or Hydrogen is far safer out of the box than Fable.

Second, single-turn retrieval (AA-LCR) no longer predicts multi-turn consistency (MLCR-AA):

Look at Gemini 4 Carbon (high) and Step 5 Preview across those columns [2]:

• On single-turn long-context retrieval (AA-LCR, 10k–100k tokens), open-weights models actually beat every proprietary flagship (Step 5 Preview 88.3%, Kimi K3.1 Pro 86.3%, MiMo-V2.6-Pro 86.3%, vs. Claude Fable 5.2 (max) 83.3%) [2].
• Once conversations span multiple turns (MLCR-AA, updated 6 Nov), Step 5 Preview drops to 18.3%, Gemini 4 Carbon (high) drops to 19.4%, and Claude Fable 5.2 (max) (43.9%), GPT-6 Astra (max) (41.7%), Claude Sonnet 5.5 (max) (41.7%), and Kimi K3.1 Pro (38.3%) pull back ahead [2]. Do not evaluate a worldbuilding model on single-turn needle-in-a-haystack scores if your design workflow runs 15–30 follow-up turns.

Third, two domain benchmarks test spatial and logical world consistency directly:

• Andon Labs Blueprint Bench 2 (spatial layout and architectural reasoning): Gemini 4 Carbon leads decisively at 0.586 normalized (0.809 raw), followed by Gemini 4 Argon (0.544 / 0.790), Claude Opus 5.5 (0.512 / 0.775), and GPT-6 Astra (0.497 / 0.768) [8]. Fable 5.2 has not been scored here yet.
• Epoch Mystery Game Puzzles (146 runs tracking clue deduction and plot-state contradictions): Claude Fable 5.2 (max) leads at 90%, ahead of GPT-6 Astra (max) (84%), GPT-6.1 Sol (max) (80%), GPT-6.1 Terra (max) (74%), Gemini 4 Carbon (high) (73%), and Claude Opus 5.5 (max) (71%), while open-weights Kimi K3.1 Pro trails at 42% [7]. That 48-point gap between Fable 5.2 and Kimi K3.1 Pro mirrors what quest designers feel when asking local models to keep multi-actor alibis consistent.

Game making, coding agents, game play and media

Beyond writing and 3D scenes, November updated coding-agent harnesses, tactical game play, and 2D/video asset generators.

Coding agents and engine arenas. Artificial Analysis materialized a fresh Coding Agents Index run on 12 November 2026, replacing October's DeepSWE v1.1 component with FrontierCode v1.1 Main alongside SWE-Atlas-QnA and Terminal-Bench v4 [9]:

Notice why the order flipped from October: Sonnet 5.5 topped October's DeepSWE-weighted composite (0.684 vs. Opus 5.5's 0.660), but under FrontierCode v1.1 Claude Opus 5.5 jumps ahead to 0.617 while Gemini 4 Carbon in Antigravity CLI edges Sonnet 5.5 (0.594 vs. 0.593) at less than two-fifths the wall time (1,790s vs. 4,785s) [9]. On budget, Codex with GPT-6.1 Sol (max) scores 0.541 at $1.50 per task, roughly one-fifth the cost of Astra ($7.04) or Carbon ($7.74) [9]. Claude Fable 5.2 is not yet part of those 8 default agent-harness rows, though in AA's base model table Fable 5.2 (max) already lifts FrontierCode 2.0 to 48.8% vs. Opus 5.5's 46.4% [2].

• On DesignArena, GPT-6 Astra (max) leads game-dev (1470 Elo, 1,089 battles) ahead of GPT-6 Astra (1460), Inkling 2 Small (california, 1433), Claude Opus 5.5 (1428), Muse Wonder 1.0 (1416), GPT-6.1 Sol (1408), Kimi K3.1 Pro (1406), and Claude Fable 5.2 (1400, early 533-battle sample) [15]. In engine-specific tracks, GPT-6 Astra (1319) and Claude Fable 5.1 (1306) top godot-game-dev, while Gemini 4 Argon (1265) leads agentic-game-dev ahead of older Claude Fable 5 (1233) [16][17]. Look at the full coding-games guide for how those one-shot browser scores translate into multi-file C# or GDScript work. (LMArena's own site blocked automated retrieval on 12 November; Epoch's latest mirrored WebDev Arena snapshot remains led by Claude Opus 5.5 at 1820.3 and GPT-6 Astra at 1800.3 [18].)

Tactical game-playing agents (SpireClimb, BALROG, GBAEval). As detailed in our game-playing AI report, general reasoning and agentic coding still fail to predict turn-by-turn strategy performance:

On both original Slay the Spire (SpireClimb) and Slay the Spire 2 (Ascension 10), OpenAI models lead by wide margins: GPT-6.1 Terra (max) ties GPT-6 Astra (max) at 34.7% on STS2 A10 while Claude Fable 5.2 (max) sits at 17.6% (and 19.6% vs. Astra's 34.0% on STS1) [10][11]. BALROG and GBAEval remain un-updated for October–November flagships, so tactical play comparisons remain incomplete outside SpireClimb [19][20].

2D concept art and video arenas. Since our game art guide, OpenAI's GPT Image 2.6 Sunburst (max) (released 4 Nov) took #1 on Artificial Analysis's Text-to-Image arena at 1261.5 Elo ($210.72 per 1,000 images, ~21 cents each), followed by GPT Image 2.5 Sunburst (max) (1229.6) and GPT Image 2.6 Flare (max) (1226.8) [21]. For high-volume sprite or UI iteration, Nano Banana 2.1 (1177.1 Elo, $33.60/1k) and ByteDance's newly released Seedream 5.5 (9 Nov, 1134.8 Elo, $20.00/1k, two cents per image) beat FLUX 3 Image (1115.1 Elo, $48.00/1k) and Qwen-Image-3.0-Pro (1096.2 Elo) on both Elo and price [21]. On AA Text-to-Video, Wan 3.0 leads at 1174.8 Elo ahead of Utopai X (1170.5), MiniMax H3 768p (1160.7), and Dreamina Seedance 2.5 (1155.3) [22].

Which model fits which job in November 2026

No single model sweeps every benchmark in November. The cleanest way to split a studio API budget is still by pipeline stage, paired with regular cost checks from our AI game cost breakdown:

How we reviewed this

We collected and cross-verified all scores directly from primary public benchmark endpoints and datasets on 12 November 2026 rather than relying on secondary blog recaps. Every model name, effort tier, price, token speed, Elo score, confidence interval, and win rate above reflects what was published in: Artificial Analysis LLM Leaderboard v5.0.1 (288 models scraped from the live page payload), AA-3D (25 models plus per-category win rates), AA Coding Agents Index (8 default agent rows materialized 2026-11-12T07:14:38Z), AA Text-to-Image and Video arenas, EQ-Bench Creative Writing v3 (145 models), DesignArena (3d-design, blender, game-dev, godot-game-dev, agentic-game-dev), BALROG, and Epoch AI's official benchmark_data.zip release (ECI, Mystery Game Puzzles, Andon Labs Blueprint Bench 2, SpireClimb STS1/STS2, Lech Mazur Writing, Fiction.liveBench, GBAEval, and mirrored WebDev Arena).

What we could not verify on 12 November: LMArena's live arena.ai web interface returned HTTP 403 to direct automated retrieval, so WebDev Arena numbers reflect Epoch AI's mirrored CSV rather than same-day live votes; Artificial Analysis reports no API pricing or speed telemetry for Muse Wonder Full (max) or Creative Writing v3 coverage for Claude Fable 5.2, Gemini 4 Carbon, and Gemini 4 Hydrogen; AA Intelligence Index v5.0.1 is rescaled and cannot be compared numerically against October's v4.3.2 numbers; and DesignArena's raw JSON uses internal codenames (california for Thinking Machines' Inkling 2 Small, grogu for Grok 4.6, and yoda for Grok 4.5). Elo scales across AA-3D, DesignArena, EQ-Bench, and AA Image/Video are independent systems and should never be compared across leaderboards.