Grok 4.5 Is Not "Opus-Class." It Is Something More Interesting: Opus-Adjacent and Much Cheaper.

Elon Musk called it directly. "It is an Opus-class model, but faster, more token-efficient and lower cost." Then, when pressed for precision, he narrowed the claim himself: "Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster."

Two different comparisons, twenty-four hours apart, from the same person, about the same model. That gap between the confident soundbite and the careful follow-up is the most honest signal in this entire launch, and it is worth reading xAI's own benchmark chart before deciding which version to believe.

Grok 4.5 launched publicly on July 8, 2026 - SpaceXAI's first model release since the company rebranded from xAI following SpaceX's February acquisition, and its first release built jointly with Cursor, the AI coding editor SpaceX agreed to acquire for $60 billion three weeks earlier. It is priced at$2 per million input tokens and $6 per million output tokens, roughly 60$5/$25. It is available today in Cursor on every plan, in Grok Build, and through the SpaceXAI API console. It is not available in the EU yet.

The capability claim and the Cursor acquisition are two separate stories that most coverage is running together. Both deserve a clear-eyed look, separately, before combining them into a verdict.

What xAI's Own Chart Actually Shows

xAI published four benchmark comparisons alongside the launch. Read exactly, not summarized: Grok 4.5 beats Claude Opus 4.8 on two of the four benchmarks the company chose to highlight - Terminal-Bench 2.1 and DeepSWE 1.0. It loses to Opus 4.8 on the other two - SWE-bench Multilingual by 6.4 points and SWE-bench Pro by 4.5 points.

Benchmark	Grok 4.5	Opus 4.8	GPT-5.5	Composer 2.5	Fable 5
Terminal-Bench 2.1	83.3%	78.9%	83.4%	73.0%	84.3%
SWE-bench Multilingual	78.0%	84.4%	77.8%	71.6%	Not reported
DeepSWE 1.0, Artificial Analysis	62.0%, high	55.8%, max	64.3%, xhigh	18.0%	66.1%, max
SWE-bench Pro	64.7%, high	69.2%, max	58.6%, xhigh	54.0%	80.3%

These are the figures shown in the official launch table. The effort settings matter: Grok 4.5 is shown at high effort on DeepSWE 1.0 and SWE-bench Pro, Opus 4.8 at max, GPT-5.5 at xhigh, and Fable 5 at max on DeepSWE 1.0.

That is a company publishing its own selected benchmark set and still landing at a 2-2 split against the model it is being marketed against. "Opus-class" is a defensible label for a model that splits four head-to-head comparisons evenly. It is a meaningfully different claim than "beats Opus," which is closer to how the soundbite spread across launch-day coverage everywhere except the outlets that actually opened the table.

Independently, Artificial Analysis places Grok 4.5 at 54 on its Intelligence Index, ranked fourth overall - ahead of every open-weight model and every Gemini model, and notably above the Chinese open-source cluster this publication covered last month in the GLM 5.2, MiniMax M3, and DeepSeek V4 comparison piece. Fourth place, at this pricing, is a genuinely strong result. It is not first place, and Musk's own walk-back to "comparable to Opus 4.7" - not 4.8 - is the more defensible version of the claim, and the independent data broadly supports that specific comparison rather than the more expansive one.

Terminal-Bench 2.1 is worth flagging separately because it is not the benchmark this publication's DeepSWE coverage found unreliable. Grok 4.5 scores 83.3% on it - a real, mechanically verifiable result on a benchmark that is harder to game through the environmental exploitation DeepSWE documented in Claude's SWE-bench Pro scores. That specific win is more trustworthy than the DeepSWE 1.0 win, precisely because Terminal-Bench tests completion of live command-line tasks rather than matching a reference implementation.

One additional wrinkle complicates the broader in-IDE performance story. Cursor's own engineering blog disclosed that an earlier Cursor codebase snapshot was accidentally included in Grok 4.5's training data - meaning the model may have had an advantage on CursorBench specifically, with the exact impact unclear. That data has been removed for future models, and Cursor is issuing a larger CursorBench update partly because of this. Treat any in-IDE performance claim as optimistic until independently reproduced.

The Number That Is Actually the Story

Buried under the "Opus-class" argument is the number that matters more for anyone actually running this in production: token efficiency.

On SWE-bench Pro, xAI reports Grok 4.5 resolves the average task using approximately 15,954 output tokens. Opus 4.8, running in its maximum-effort mode, uses 67,020 tokens for the same benchmark - a 4.2x gap. Combined with the per-token pricing, that efficiency claim is arguably more interesting than the raw capability comparison, because it is the number that actually determines what a model costs to run at scale, and it is harder to inflate through benchmark-specific tricks than a pass/fail score is.

One independent reviewer ran a full battery of build-and-iterate tasks through Grok Build - several 3D scenes, a Linux driver task, a 3D-printable engine model, sprite-repair iteration, and a C++ game with follow-up fixes - and the total API cost across roughly a dozen distinct tasks came to $7.78. That is not "essentially free," and it should not be read as a controlled benchmark. But as a real-world reference point for what a working session actually costs, it is a useful data point that the benchmark charts do not provide.

The unanswered question, which the same reviewer named directly, is reliability under messier conditions than a curated benchmark: whether the token efficiency holds up across inconsistent repos, weird enterprise documents, and long-running agent loops, or whether the savings on the API bill get eaten by additional review time when the model cuts corners to hit its efficient output. Benchmarks measure capability. They do not measure how much human oversight a model requires to trust its output in production, and that gap is exactly where the DeepSWE findings on Claude's git-history exploitation and METR's reward-hacking finding on GPT-5.6 both landed earlier this month. Grok 4.5 has not yet had an equivalent independent audit.

Why This Model Exists: The Cursor Acquisition

Understanding Grok 4.5 requires understanding the deal that produced it, and the deal is a genuinely strange story even by 2026's standards.

On April 21, 2026, SpaceX secured an option to acquire Anysphere - Cursor's parent company - for $60 billion later in the year, or alternatively pay roughly$10 billion for a partnership. If no deal closed, SpaceX had agreed to pay Cursor a $1.5 billion termination fee plus$8.5 billion in computing resources regardless. This was not an impulsive purchase. It was structured months in advance, with SpaceX holding an option no other bidder could match.

On June 16, five days after SpaceX's IPO priced, SpaceX exercised that option. The all-stock deal, structured as a reverse triangular merger, is expected to close in Q3 2026. SpaceX shares jumped roughly 16% on the announcement, briefly pushing the company's market cap to $2.94 trillion and past Microsoft and Amazon.

The reason Cursor sold rather than continuing to operate independently is the part of the story most coverage skips. Cursor reached roughly $4 billion in annualized revenue in under four years - one of the fastest ARR ramps in software history - but its market share, per Ramp's corporate spending data, fell from approximately 41$60 billion look like rescue rather than retreat, whatever the announcement language said about partnership.

Both OpenAI and Microsoft had pursued Cursor before SpaceX. OpenAI made two separate approaches; Cursor rebuffed both, reportedly to preserve independence. Microsoft examined a bid and passed. SpaceX won because it had already locked in an option nobody else could match, executed within days of having fresh IPO stock to pay with.

What Neutrality Actually Meant, and What Its Loss Means

Cursor's entire enterprise pitch for three years was neutrality: pick Claude for reasoning-heavy work, GPT for breadth, Gemini for vision, switch freely, no lock-in to a single lab's roadmap or pricing. Over 50,000 enterprise clients and developers at roughly two-thirds of the Fortune 500 adopted Cursor specifically because it sat above the frontier labs rather than inside one of them.

That neutrality is now structurally compromised, not by any announced policy change, but by ownership. SpaceXAI competes directly with Anthropic and OpenAI on the model layer. Maintaining genuinely first-class support inside Cursor for two direct competitors' models, while the parent company's own model needs the developer mindshare and training signal that Cursor's user base represents, is a tension that does not require a single explicit decision to resolve unfavorably over time. It resolves through which model gets the best default settings, which gets the fastest new-feature integration, which gets prioritized when infrastructure is constrained.

There is a direct precedent for how this plays out under pressure rather than goodwill. When OpenAI was acquiring Windsurf, another AI coding startup, Anthropic cut off Windsurf's Claude access. The stated reason was capacity allocation. The practical effect was that an OpenAI-acquired coding product lost access to a competitor lab's model entirely. Anthropic and OpenAI both retain that same commercial leverage over Cursor now that it sits inside a direct competitor's ownership structure. Nothing requires either lab to exercise it. The Windsurf case establishes that they have, when the ownership alignment made it strategically obvious to do so.

Mitch Ashley of The Futurum Group named the actual shift in incentive precisely: enterprise buyers now have to re-underwrite Cursor as a single-owner dependency, where the operative question is no longer just what the tool does, but whether a newly public aerospace-and-AI conglomerate with no enterprise software track record will maintain a developer layer whose model partners it now competes against.

Developer reaction has split along an almost perfectly predictable line: people who have never used Cursor are skeptical of the drama; people who use it daily are anxious, and the anxiety has a specific, well-documented history behind it. Atom was sunsetted after Microsoft's GitHub acquisition. Parse shut down after Facebook acquired it. Developer tools under new corporate ownership do not have a reassuring track record, even when the acquiring company makes public commitments to preserve the product.

Where Grok 4.5 Actually Fits

Setting the marketing language aside, the honest positioning is this: Grok 4.5 is a real, fourth-place frontier-adjacent model with genuinely aggressive pricing and a genuinely strong token-efficiency story, shipped as the first visible output of a $60 billion vertical-integration bet whose real strategic purpose is not the model itself but the developer distribution and training-data pipeline that owning Cursor provides.

It is not "Opus-class" in the sense of beating Opus 4.8 outright - xAI's own numbers do not support that framing. It is "Opus-adjacent" in the more precise, more defensible sense Musk retreated to within a day of the launch: comparable to Opus 4.7, meaningfully faster, and available at a fraction of the cost, with the caveat that some in-IDE performance claims may be partially inflated by an accidental training-data advantage on CursorBench, the benchmark most relevant to its primary distribution channel.

For a developer already inside Cursor, trying Grok 4.5 costs nothing beyond a model switch, and the $7.78 real-world session cost is a genuinely useful data point regardless of how the benchmark disputes resolve. For an enterprise evaluating whether to keep standardizing on Cursor at all, the model's benchmark performance is now the less important variable. The more important one is whether a tool whose entire value proposition was neutrality between AI labs can remain neutral once it is owned by one of them - and the Windsurf precedent suggests that question does not stay hypothetical for long once the strategic incentives point in one direction.

Sources:

• Introducing Grok 4.5 - SpaceXAI
• Introducing Grok 4.5 - Cursor
• Grok 4.5 brings SpaceXAI to the intelligence frontier - Artificial Analysis
• SpaceX filing on the Cursor merger - US Securities and Exchange Commission

Previously on TheQuery: The Real AI Race Nobody Is Covering: MiniMax M3, GLM 5.2, and DeepSeek V4 Are Fighting Each Other and GPT-5.6 Is Now Publicly Available