Grok 4.5 is HERE! (Real Tests and Review)

Grok 4.5 lands fast and cheap at $2.33 for four mini projects, but quality splits 50/50 against Fable 5 and GPT 5.5.

Specs and context

Grok 4.5 ships with a 500k context window (text + image), with Elon Musk saying a 1M upgrade is coming "by next week." The 500k window still carries the same regression xAI has had on prior Grok models, which is part of why long-context output can drift. Benchmarks position it slightly below GPT 5.5 on SWE-bench Verified 1.0, above GLM 5.2 on SWE-bench Verified 1.1, and best-in-class on SWE Marathon (autonomous long-horizon coding).

Coding test results

Across four projects — 3D Chinese timber architecture, medieval castle, Helm's Deep, and Hogwarts — Grok 4.5 cost $2.33 total. The Hogwarts build was "way too bright" and uninteractive, and Helm's Deep was "way too dark… big fail." The Chinese architecture build on first try was a pagoda with no real roof; after the creator stripped the photo-upload feature, the second pass produced a pillar-erection animation similar to Fable 5, but the "age erosion" is still a flat color change with no cracked-wood texture that Fable 5 nails.

Where it actually wins

Elon's own internal assessment is that Grok 4.5 is "roughly comparable to Opus 4.7." On a real code review of the creator's smolnews X-news pipeline (built earlier with Fable 5 plan + MiniMax M3 execution), Grok 4.5 caught the actual gap: the agent's "time window consistency" when filtering relevant vs. irrelevant tweets. If your pipeline touches X data, scraping, or news aggregation, Grok 4.5 is the right reviewer.

Bottom line

It is "definitely nowhere close to Fable 5" in raw strength, but at $2.33 vs. $30+ for the same four projects, the cost gap is the only thing most people will feel.