grok 4.5 vs fable 5 vs gpt 5.5 vs glm 5.2

today @SpaceXAI released grok 4.5 – a model built around coding, agentic tasks and knowledge work. key facts:

• trained across tens of thousands of nvidia gb300 gpus, with rl over hundreds of thousands of tasks centered on multi-step software engineering

• served at 80 tps, priced at $2/m input and $6/m output

• xai claims ~2x token efficiency: 15,954 avg output tokens per swe bench pro task vs 67,020 for opus 4.8 max

• published benchmarks: swe marathon 29.0% (1st), terminal bench 2.1 83.3% (3rd), deepswe 1.0 62.0% (3rd), swe bench pro 64.7% (3rd) – fable 5 leads 3 of the 4

our test – 3 prompts, single-file html, three.js:

1. photorealistic 3d interior planner for a japanese-style bedroom
2. 3d interior planner for a 1980s memphis-design home office, flat cel-shaded – not photoreal
3. architectural 3d interior planner for a loft kitchen, rendered as a live technical blueprint

results:

- cost
#1 grok 4.5 – $0.32
#2 glm 5.2 – $2.46
#3 gpt 5.5 – $2.63
#4 fable 5 – $7.20

- output tokens
#1 grok 4.5 – 41,368
#2 gpt 5.5 – 65,940
#3 fable 5 – 110,368
#4 glm 5.2 – 401,698

- attempts
#1 grok 4.5 – 3
#1 fable 5 – 3
#2 glm 5.2 – 4
#3 gpt 5.5 – 5

- lines of code
#1 fable 5 – 1,535
#2 gpt 5.5 – 2,429
#3 glm 5.2 – 2,835
#4 grok 4.5 – 3,318

observations:

• grok 4.5 takes cost and token count outright and ties fable 5 on attempts, but writes the most code to get there

• fable 5 is 22x more expensive per run and produces the leanest files – 1,535 loc against grok's 3,318.

• glm 5.2 spent 401k tokens, nearly 10x grok, for the third-longest output. the token efficiency claim holds up

grok's code quality – reviewed by opus 4.8:

upsides: all three run first try, zero external assets. every texture is procedural – tatami weave, wood grain, plaster, terrazzo, animated crt scanlines. rotation-aware aabb collision, 900mm clearance-gap detection, kelvin-to-rgb sun model tied to tone-mapping exposure

downsides: three.rgbformat in the memphis file – removed in r137, file pins r160, cel-shading bands silently break. dead code throughout: an unused layouts array, mat positions set twice, three overlapping loops in getmovablefromintersect where two are unreachable. new materials allocated on every mode switch. no disposal anywhere

pattern: dense, working, visually ambitious code, fast and cheap – and no cleanup. it optimizes for a demo that runs, not code that gets reviewed