Musk Unveils Grok 4.5: First Benchmarks Reveal Its Strengths and Limits

Benchmark performance

Terminal‑Bench 2.1 83.3 % – matches GPT‑5.5 (83.4 %) and is 1 point below Fable 5 (84.3 %).

SWE‑Bench Multilingual 78.0 % – slightly above GPT‑5.5 (77.8 %) but below Opus 4.8 (84.4 %).

DeepSWE 1.0 62.0 % – lower than GPT‑5.5 (64.3 %) and Fable 5 (66.1 %) but higher than Opus 4.8 (55.8 %).

SWE‑Bench Pro 64.7 % – better than GPT‑5.5 (58.6 %) but behind Opus 4.8 (69.2 %) and far behind Fable 5 (80.3 %).

Pricing and inference speed

Input token price $2/​million tokens, output token price $6/​million tokens. Inference throughput reaches up to 80 TPS , claimed to be faster than typical flash models.

Third‑party evaluations

CursorBench places Grok 4.5 third in the High‑score tier, 4 % behind Fable 5 while costing roughly one‑tenth of Fable 5.

Artificial Analysis gives Grok 4.5 a total score of 54 (fourth place). Its Codex sub‑score is comparable to GPT‑5.5 but achieved at a much lower cost.

Qualitative demos

Visual demos include a Minecraft‑style pixel‑art game generated from a single prompt, a clock‑face rendering, and particle‑effect animation. The “lava‑lamp” BridgeBench test, which measures UI aesthetics and visual design, shows Grok 4.5 trailing the competitors.

Cost comparison

Grok 4.5 – $2 input / $6 output GPT 5.6 – $5 input / $30 output Opus 4.8 –

$5 input / $25 output

Improvement over Grok 4.3

Performance gains are evident across most metrics compared with the previous Grok 4.3 release, though Grok 4.5 still does not fully match Fable 5 on the hardest engineering benchmarks.

Images