Key Takeaways

• Grok 4.5 is genuinely good but not the best. It scores 54 on the Artificial Analysis Intelligence Index, placing fourth behind Claude Fable 5, GPT-5.5 and Claude Opus 4.8.
• Released on July 8, 2026, it is SpaceXAI’s first model built specifically for coding and agentic work, trained partly on real Cursor developer sessions.
• Its standout strength is cost efficiency: roughly $0.31 per Intelligence Index task, and 1.9 million tokens per coding task versus 6–7 million for its priciest rivals.
• It leads outright on the SWE Marathon coding benchmark (29% pass-at-one) and edges GPT-5.5 on the τ³-Banking agentic test.
• It trails badly on the hardest long-horizon coding evaluation, SWE-Bench Pro (64.7% versus Fable 5’s 80.4%).
• A real weakness: its hallucination rate more than doubled from 25% to 54% versus Grok 4.3, even as raw accuracy improved.
• xAI has not published a Grok 4.5 model card, unlike Anthropic and OpenAI, which ship safety documentation at launch.

Grok 4.5 is a strong, near-frontier model that wins on price and speed rather than raw intelligence. On independent testing from Artificial Analysis, it lands fourth on the Intelligence Index at a score of 54, behind Claude Fable 5, GPT-5.5 and Claude Opus 4.8, but it reaches that level of quality at roughly a fifth of their per-task cost. So the honest answer to “how good is it” is that Grok 4.5 is very good at well-scoped, checkable work, and less dependable on the hardest problems where accuracy matters most.

SpaceXAI, the merged SpaceX and xAI entity that most people still call xAI, released the model on July 8, 2026 as its first release built for coding and agentic tasks. Elon Musk described it as an “Opus-class model, but faster, more token-efficient and lower cost.” The independent numbers mostly support the speed and cost half of that claim, while the “Opus-class” label oversells the intelligence half.

What Grok 4.5 Actually Is

Grok 4.5 is built on a foundation model xAI calls V9, reported at roughly 1.5 trillion parameters and trained on the company’s Colossus cluster using tens of thousands of Nvidia GB300 GPUs. It supports a 500,000-token context window, accepts text and images, and returns text only. Reasoning effort is configurable across low, medium and high, with high the default and the setting used for the benchmark scores below. The model’s knowledge cutoff is February 1, 2026.

The training strategy is the interesting part. xAI trained Grok 4.5 alongside Cursor, folding real IDE session data from Cursor users into its post-training for coding and agentic behaviour. That is a meaningfully different approach from chasing abstract reasoning puzzles, and it shows up in where the model wins: agentic coding in context, not exam-style problems.

The Benchmarks: Where Grok 4.5 Wins and Where It Loses

On coding agents, Grok 4.5 is competitive. It scores 76 on the Coding Agent Index, level with GPT-5.5 running in Codex and one point behind Fable 5. On Terminal-Bench 2.1 it reaches 83.3%, within a point of both Fable 5 (84.3%) and GPT-5.5 (83.4%), and clearly ahead of Opus 4.8. It takes the top spot on SWE Marathon, a single-attempt measure of resolving realistic engineering tasks, at 29% against Opus 4.8’s 26%.

The picture darkens on the hardest tests. On DeepSWE 1.1 it scores 53%, well behind GPT-5.5’s 67% and Fable 5’s 70%. On SWE-Bench Pro, a tougher and less-contaminated benchmark, it manages 64.7% against Opus 4.8’s 69.2% and Fable 5’s 80.4% — the widest deficit in the field. These are the tasks that demand sustained reasoning over large codebases, and Grok 4.5’s efficiency-first design shows its ceiling there. For readers weighing it against Anthropic’s line, our comparison of Claude Opus 4.6 vs 4.7 vs 4.8 puts those coding numbers in context.

Benchmark	Grok 4.5	Claude Opus 4.8	Claude Fable 5
Intelligence Index (Artificial Analysis)	54 (4th)	Higher	Top
Terminal-Bench 2.1	83.3%	78.9%	84.3%
SWE Marathon (pass-at-one)	29%	26%	24%
SWE-Bench Pro	64.7%	69.2%	80.4%
Cost per Intelligence Index task	$0.31	$1.80	$2.75

The Honesty Problem Nobody Put on a Slide

The most important number xAI left out of its launch materials concerns calibration. On the AA-Omniscience Index, which tests whether a model knows what it does not know, Grok 4.5’s accuracy improved from 35% to 52%. But its hallucination rate rose faster, from 25% to 54%. In plain terms, the model now holds more knowledge yet states false things with more confidence than its own predecessor did.

That pairing matters for anyone running unsupervised or high-stakes work. A model that rarely says “I don’t know” is riskier in production than one that is simply less capable. Research assistance, customer-facing agents and financial or legal summarisation are exactly the settings where teams should test Grok 4.5 against their own material before trusting it at volume.

Speed, Efficiency and Access

Grok 4.5 generates around 119 tokens per second with a time-to-first-token near 13.7 seconds. It is not the fastest model available, but because it burns so few tokens per task, it finishes agentic work faster and cheaper than models that stream tokens quicker yet need far more of them. It is available through Grok Build, inside Cursor on every plan, via the xAI console API, and through Microsoft Office plugins for Word, PowerPoint and Excel. European availability arrived shortly after the mid-July rollout.

This efficiency-first identity fits the wider direction xAI has taken since folding into SpaceX. The SpaceX and xAI merger tied Grok’s future to cheap compute and energy at scale rather than to winning the next reasoning benchmark, and Grok 4.5 is the first model that plainly carries that logic. Where it sits against its main rivals is covered in our look at xAI versus OpenAI versus Anthropic.

So, Is Grok 4.5 Good? The Verdict

Grok 4.5 is a highly capable, highly efficient workhorse. For high-volume coding agents, routine code review, customer-support automation and data-processing pipelines where a human or an automated check can catch the occasional wrong answer, it is one of the best value-for-money choices available. For unsupervised, high-stakes tasks that must be right the first time — complex refactors, novel research, regulated summarisation — Opus 4.8, Fable 5 or GPT-5.5 remain the safer picks, and the missing model card gives regulated industries an extra reason to wait. Teams comparing options for day-to-day building will also find our tier list of the best LLMs for vibe coding a useful cross-check.

Model benchmarks are anchored to Artificial Analysis testing published on July 9–10, 2026, and to launch reporting from mid-July 2026. Figures move as new evaluations arrive, and this is informational rather than investment or procurement advice.

If you are interested in this topic, we suggest you check our articles:

• ChatGPT-5 vs Grok 4: Who Wins?
• SpaceX-xAI Merger: What It Means for Grok and AI Output
• xAI vs OpenAI vs Anthropic: Which AI Lab Wins in 2026?
• Best AI LLM for Vibe Coding: Complete 2026 Tier List
• 2026 AI Subscription Prices: Gemini vs ChatGPT vs Claude

Sources: AI Tools Review, The Decoder, DataCamp, Fello AI

Written by Alius Noreika