Grok 4.5 Review: Opus-Class Coding at a Fraction of the Cost

A first-look review based on SpaceXAI's launch benchmarks, the official Grok 4.5 model page, and early third-party coverage — not extended hands-on use. We'll revisit it after living with the model.

The Bottom Line (July 2026)

Grok 4.5 is the best value in frontier coding right now. It isn't the single smartest model — Fable 5 and OpenAI's GPT-5.5 still edge it on most benchmarks — but it delivers roughly Opus-class results while burning about a quarter of the tokens and charging a third of the price. For anyone running coding agents for hours and paying per token, that trade is more compelling than another point on a leaderboard. Released July 16 and trained alongside Cursor, it's free for a limited time in Grok Build and Cursor, which makes trying it a no-brainer.

Performance: top-tier, not the crown

On raw capability, Grok 4.5 sits firmly in the top tier without leading it. It posts 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-Bench Pro — a hair behind Fable 5 and GPT-5.5 on both. Where it genuinely wins 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 (26.0%) and Fable 5 (24.0%). That's the tell: Grok 4.5 is built to grind through long, real engineering sessions rather than to win a one-shot prompt.

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 leaderboards, as reported by xAI. A dash means no comparable figure was published.

Efficiency: the real headline

The number that matters most isn't a benchmark score — it's tokens. Grok 4.5 resolves a SWE-Bench Pro task in about 15,954 output tokens on average, roughly 4.2× fewer than Opus 4.8's ~67,020, and it's served at fast-model speeds of 80 tokens per second. Pair that with $2 per million input and $6 per million output — less than half the price of the leading closed models — and the cost per finished task collapses. xAI's slogan for it, "the highest intelligence per unit of time and cost," is marketing, but the underlying math holds up.

Ecosystem and caveats

The catch is lock-in and governance. Grok 4.5 is at its best inside the SpaceX family — it's the default in Grok Build and it trained on real Cursor sessions, both now under the same corporate roof after Elon Musk's acquisition spree. That vertical integration is a genuine data-flywheel advantage, but it also means Grok's strongest experience lives inside xAI's own surfaces. Enterprises will also want to weigh data-governance questions that follow anything under Musk's control, and note that the current free access is temporary before it shifts to usage pricing. Outside coding, agentic work, and Office document generation, Grok 4.5's strengths are less proven.

Verdict: 4.4 / 5

Grok 4.5 doesn't dethrone the frontier — but it reframes the question. If you judge a coding model by its single best benchmark, buy Fable 5 or GPT-5.5. If you judge it by how much good work it ships per dollar and per minute, Grok 4.5 is arguably the smartest purchase on the board today. It's a value play executed with real conviction, and while it's free in Grok Build and Cursor, there's no reason not to put it to work.