SpaceXAI released its flagship AI model, Grok 4.5, on July 8, 2026. Elon Musk described it as an "Opus-class model," tuned for coding, agentic tasks and knowledge work while running faster and at lower cost than earlier versions.
July 8, 2026 · SpaceXAI
Grok 4.5 bets on cost efficiency over benchmark bragging rights
An "Opus-class" model built on xAI's V9 foundation, tuned for coding and real-world engineering — pitched as faster, more token-efficient and cheaper than the top-tier rivals it targets.
~1.5T
V9 foundation model parameters
500k
Context window (tokens)
$2 / $6
API input / output per 1M tokens
The edge: token efficiency on SWE Bench Pro
Average output tokens per task — fewer is cheaper. Grok 4.5 uses roughly 4.2× fewer than Opus 4.8.
≈ 4.2× fewer output tokens per task
Terminal Bench 2.1 — vendor-stated scores
Benchmarks are vendor-stated and await independent verification. On DeepSWE 1.0 and SWE Bench Pro raw scores, Grok 4.5 trails Fable 5 in some cases.
The "data flywheel"
Trained on the Colossus cluster of tens of thousands of NVIDIA GB300 GPUs, with usage feeding back into retraining.
Model uses coding tools
→
Cursor usage data collected
→
Large-scale RL retraining
Plan: a new from-scratch foundation model every month through 2026.
Where it wins
One-shot website and app builds
Practical repository analysis
Real-world manufacturing productivity, valued over benchmark scores
Token efficiency + low cost
Open questions
Raw benchmark scores trail Fable 5 in some cases
Figures are vendor-stated, unverified
Not available in the EU until mid-July
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…