Reviews · JULY 11, 2026
Grok 4.5 lands at $2/$6, trades a little SWE-Bench Pro headroom for 4.2× fewer output tokens
SpaceXAI's first post-merger flagship, co-trained with Cursor on tens of thousands of GB300s, ships at a fifth of Opus 4.8's output price and resolves SWE-Bench Pro tasks in 15,954 tokens against Opus 4.8's 67,020.
SpaceXAI shipped Grok 4.5 on July 8, 2026 at $2 per million input tokens and $6 per million output tokens, and the price is doing more analytical work than the benchmark chart. Anthropic's Opus 4.8 sits at $5 in and $25 out. OpenAI's GPT-5.6 Luna, reaching general availability the following day, matches Grok's $6 output but at $1 input. The Grok output price is a fifth of Opus 4.8's.
On raw capability, Grok 4.5 is a close second-tier finisher. It resolves 64.7% of SWE-Bench Pro tasks against 69.2% for Opus 4.8 (max) and 64.3% for Opus 4.7, a roughly 4.5-point deficit against the current Anthropic flagship. Fable at max configuration takes the top slot at 80.4%. On Terminal-Bench 2.1, Grok 4.5's 83.3% trails GPT-5.5 at 83.4% and Fable at 84.3% by rounding-error margins. It leads SWE Marathon pass@1 at 29.0% over Opus 4.8's 26.0% and Fable's 24.0%, and lags on DeepSWE 1.0 at 62.0% behind Fable's 66.1% and GPT-5.5's 64.31%.
The interesting number is elsewhere on the card. Grok 4.5 resolves the average SWE-Bench Pro task in 15,954 output tokens. Opus 4.8 (max) needs 67,020. That's a 4.2× token-efficiency gap, stacked on top of a per-token output cost that's more than 4× cheaper, which puts the price-per-resolved-task delta at roughly an order of magnitude. For any agentic workload that pays per token and runs at scale, that's the review.
The provenance matters. This is the first flagship since xAI folded into SpaceX, co-trained with Cursor, which SpaceX acquired for $60 billion in a deal that closed weeks earlier. "We've partnered with SpaceXAI to train Grok 4.5," Cursor said. Training ran across "tens of thousands" of NVIDIA GB300 GPUs, capacity Axios notes SpaceXAI is currently leasing to Anthropic and Google. Elon Musk called it "an Opus-class model, but faster, more token-efficient and lower cost" on X, and internally described its capabilities as "roughly comparable to Opus 4.7." The benchmark spread confirms both readings.
Musk's framing is unusually disciplined by his standards, and it maps to a specific strategic bet. The 2017 GPU-compute cycle rewarded whoever posted the top MMLU number. This cycle, with agentic harnesses running unattended and enterprise finance and legal deployments billing by the million tokens, rewards whoever posts the lowest cost-per-resolved-task. Grok 4.5 isn't the best model on any leaderboard the vendor published. It's the cheapest way to finish the job on most of them.
EU availability is deferred to mid-July, with no regulatory reason offered. The silence is its own signal.
Sources
- https://x.ai/news/grok-4-5
- https://money.usnews.com/investing/news/articles/2026-07-08/spacexai-launches-grok-4-5-model-for-coding-agentic-tasks
- https://techcrunch.com/2026/07/08/spacexai-releases-grok-4-5-which-elon-describes-as-an-opus-class-model/
- https://www.bloomberg.com/news/articles/2026-07-08/spacexai-cursor-unveil-grok-ai-model-for-legal-finance-tasks
- https://www.axios.com/2026/07/08/spacexai-grok-new-model