SpaceXAI's Grok 4.5 Bets on Efficiency Over the Crown — Opus-Class Coding at 4.2× Fewer Tokens

SpaceXAI — the merged SpaceX–xAI — has launched Grok 4.5, its "smartest model ever" and its first flagship built specifically for coding, agentic tasks, and knowledge work. Released on July 16 and trained alongside Cursor, the model is available today in Grok Build, in Cursor on every plan, and through the SpaceXAI API. But the more interesting story isn't that Grok 4.5 wins — it's that xAI has stopped trying to win the way everyone else does.

On the raw benchmarks, Grok 4.5 is very good but not the champion. xAI's own charts show Fable 5 (max) and OpenAI's GPT-5.5 (xhigh) edging it out on DeepSWE and Terminal-Bench, where Grok posts a strong 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-Bench Pro. Where it does lead is SWE Marathon, a test of sustained multi-hour agentic work, resolving 29.0% of tasks pass@1 — ahead of Anthropic's Opus 4.8 at 26.0% and Fable 5 at 24.0%. In other words, Grok 4.5 is roughly Opus-class: not the single smartest model on the board, but firmly in the top tier.

Benchmark	Grok 4.5	Fable 5 (max)	GPT-5.5 (xhigh)	Opus 4.8 (max)
Terminal-Bench 2.1	83.3%	84.3%	83.4%	78.9%
SWE-Bench Pro (resolve)	64.7%	80.4%	58.6%	69.2%
DeepSWE 1.0 (pass@1)	62.0%	66.1%	64.3%	55.8%
SWE Marathon (pass@1)	29.0%	24.0%	—	26.0%
Output tokens / task	~15.9k	—	—	~67k
Price ($ / 1M in–out)	$2 / $6	—	—	—

Competitor figures are drawn from each developer's published system cards or benchmark leaderboards, as reported by xAI. A dash means no comparable figure was published.

The pitch is what happens when you divide that intelligence by time and money. On the same SWE-Bench Pro tasks, Grok 4.5 resolves each one in about 15,954 output tokens on average — roughly 4.2× fewer than Opus 4.8's ~67,020 — while being served at fast-model speeds of 80 tokens per second. Priced at $2 per million input tokens and $6 per million output, it undercuts the leading closed models by more than half, and the token efficiency compounds that: fewer steps, fewer tokens, lower bill. xAI's claim is blunt — "the highest intelligence per unit of time and cost."

That efficiency is the product of a very deliberate — and very vertically integrated — training pipeline. Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning scaled across hundreds of thousands of tasks centered on multi-step software engineering, and an asynchronous stack that lets agentic rollouts run for hours while learning continues. Crucially, it was trained alongside Cursor — which, like xAI, now sits inside SpaceX after Elon Musk's acquisition spree — giving the model a real-world coding data flywheel its rivals have to license or scrape. Beyond code, Grok 4.5 is the new default in Grok Build, where it builds multi-sheet Excel models, native-shape PowerPoint diagrams, and Word documents, positioning it as an office agent as much as a coding one.

Strategically, Grok 4.5 reframes the race. As frontier labs chase the top of the leaderboard with ever-larger, ever-pricier models, xAI is competing on the economics of agentic work — where an assistant runs for hours and every token has a cost. By pairing Opus-class quality with a fraction of the tokens, fast serving, low prices, and a captive distribution channel in Cursor and Grok Build, SpaceXAI is betting that for developers actually shipping software with AI, the cheapest good-enough model wins more often than the single best one. To seed that bet, Grok 4.5 is free for a limited time in both Grok Build and Cursor.