On July 8, 2026, Elon Musk, SpaceXAI's founder, announced that his company was ready to let the public judge its newest model.
"It is an Opus-class model, but faster, more token-efficient and lower cost." — Elon Musk, on X, July 8, 2026
He was comparing Grok 4.5, the newest release from SpaceXAI, directly to Anthropic's flagship line. In a follow-up post the same day, he narrowed the claim to a specific number: Grok 4.5, he wrote, was "roughly comparable to Opus 4.7, but much faster."
The independent scoreboard told a different story.
Artificial Analysis, the benchmarking firm that ML and AI engineering teams increasingly treat as a neutral referee, ran Grok 4.5 through its Intelligence Index within a day of launch. The model scored 54, placing it fourth: behind Claude Fable 5, OpenAI's GPT-5.5, and Claude Opus 4.8, the exact model Musk had just claimed to match.
For engineers deciding where to route agentic coding work this week, the gap between the marketing and the measurement matters more than usual. Grok 4.5 is aggressively cheap, which is real and independently confirmed. But the same 24 hours of testing also surfaced a hallucination rate that more than doubled from the previous Grok generation, plus a benchmark that SpaceXAI's own coding partner had to withdraw from the launch after discovering the model had trained on a snapshot of that partner's own code.
Here is what the independent numbers actually show, and what SpaceXAI left out of its launch materials.
Fourth Place, Not First
Grok 4.5 is not a weak model. On the Artificial Analysis Intelligence Index, a composite built from nine separate evaluations spanning coding, reasoning, agentic work, and general knowledge, it improved 16 points over its predecessor, Grok 4.3. Artificial Analysis said that puts SpaceXAI behind only OpenAI and Anthropic, ahead of every open-weights model on the board and ahead of Google's Gemini line as well.
On specific agentic tasks, Grok 4.5 topped the field outright:
- 𝜏³-Banking (agentic tool use): Grok 4.5 scored 33%, the best of any model Artificial Analysis tracked, ahead of GPT-5.5's 31%.
- SWE Marathon (long-horizon software engineering): Grok 4.5 finished first at 29%, ahead of Opus 4.8's 26% and Fable 5's 24%.
- GDPval-AA v2 (real-world agentic knowledge work): Grok 4.5 posted an Elo rating of 1543, between Opus 4.8's 1600 and GLM-5.2's 1513.
None of that changes the headline number. A score of 54 is a real result, not a talking point. It is also fourth place on a leaderboard where Musk had just claimed a tie for first.
The Road to Launch Day
Grok 4.5 did not appear out of nowhere. It is the product of a year of corporate consolidation that reshaped what xAI even means.
Cursor engineers had reportedly been working alongside SpaceXAI staff for weeks before the deal formally closed, according to Bloomberg's reporting on the acquisition. That closeness between the two companies did not start with the acquisition; LDS has previously reported on Musk's early moves to bring Cursor engineers into xAI's orbit. By the time Grok 4.5 launched, the two companies were not really separate anymore. That closeness is exactly what produced the story's sharpest complication.
A Benchmark Trained on the Answer Key
Cursor and SpaceXAI describe Grok 4.5 as a joint effort. Cursor's own launch post says the model trained on trillions of tokens of Cursor data: real debugging sessions, multi-file edits, and developer-agent interactions pulled from its product.
That arrangement produced a specific problem. Buried in a footnote on Cursor's own blog post, the company disclosed what happened with CursorBench, its internal evaluation built from coding tasks tied to its own codebase.
"Grok 4.5 has an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training. The exact impact is unclear. That data has been removed for future models, and in parallel we are working on a larger update to CursorBench, hence the exclusion here." — Cursor, company blog, July 8, 2026
In plain terms: the model that Cursor's own benchmark was supposed to test had, at some point in training, already seen a version of the material the test was built from. Cursor pulled the score rather than publish a number it could not vouch for.
The timing makes the disclosure harder to wave off as routine data hygiene. SpaceX did not just partner with Cursor. It bought the company outright for $60 billion, closing the deal less than a month before Grok 4.5 shipped. A model trained jointly with a subsidiary, evaluated on that subsidiary's own benchmark, that then has to be pulled because the model had already seen the answers, is close to a textbook conflict of interest. Third-party evaluations built on independent infrastructure, including DeepSWE and SWE-Bench Pro, were not affected and remain intact.
The Hallucination Rate That More Than Doubled
The most consequential number for anyone using Grok 4.5 on factual or client-facing work is not the Intelligence Index score. It is the model's performance on AA-Omniscience, Artificial Analysis's index for knowledge reliability, which rewards correct answers, penalizes confident wrong ones, and does not penalize a model for declining to answer.
Grok 4.5 scored 26 on AA-Omniscience, up from Grok 4.3's 18. That gain came from real improvement: raw accuracy rose from 35% to 52%.
But the hallucination rate, the share of wrong answers delivered with confidence, rose even faster: from 25% to 54%.
Artificial Analysis described this as a familiar pattern. Larger models tend to know more while also growing more confident about the answers they get wrong. Grok 4.5 fits that pattern closely: it is more likely to be right than its predecessor, and when it is wrong, it is now more likely to sound certain about it.
A separate benchmark reinforces the same concern from a different angle. On AutomationBench-AA, an Artificial Analysis test that scores whether an agent can complete real workplace tasks across simulated Gmail, Slack, Salesforce, and HubSpot environments without breaking rules, Grok 4.5 actually finished first, completing 51.4% of tasks cleanly. But it also logged 0.63 guardrail violations per task, the worst rate among the models Artificial Analysis compared it against, ahead of Opus 4.8's 0.55 and Google's Gemini 3.5 Flash at 0.46. The model that won on completion also broke the most rules doing it.
What a Neutral Referee Shows
SpaceXAI's own launch chart included a detail that undercuts its headline framing, if you read the fine print.
The chart shows Grok 4.5 on two versions of the DeepSWE benchmark. On DeepSWE 1.0, evaluated using each model provider's own coding harness, Grok 4.5 scored 62%, a reasonably strong third place behind Fable 5 and GPT-5.5. On DeepSWE 1.1, evaluated with a single standardized agent that Datacurve controls rather than each lab's own tooling, Grok 4.5's score dropped to 53%, trailing GPT-5.5's 67% and Fable 5's 70%, and falling behind Opus 4.8 as well.
The gap between those two numbers is the gap between a lab grading its own work and a neutral party running the same test the same way for everyone. It is not unique to SpaceXAI; OpenAI faced a similar credibility question this year over how its own coding benchmarks were scored, a pattern LDS examined in its reporting on GPT-5.6 Sol's disputed coding record. A chart published by the company selling the model is not the same evidence as a chart published by the company grading it.
The Other Side: Why the Price Might Still Win
None of this means Grok 4.5 is a bad choice for every team. The pricing argument is real, and it is not marketing spin.
| Model | AA Intelligence Index Rank | Input Price (per 1M tokens) | Output Price (per 1M tokens) | Coding-Agent Cost per Task |
|---|---|---|---|---|
| Claude Fable 5 | #1 | $10 | $50 | $11.80 (in Claude Code) |
| GPT-5.5 | #2 | $5 | $30 | $5.07 (in Codex) |
| Claude Opus 4.8 | #3 | $5 | $25 | Not measured on this index |
| Grok 4.5 | #4 (score: 54) | $2 | $6 | $2.49 (in Grok Build) |
Grok 4.5 also resolves the average SWE-Bench Pro task using 15,954 output tokens, against Opus 4.8's 67,020, a 4.2 times efficiency gain that Artificial Analysis independently confirmed rather than simply repeating from SpaceXAI's marketing. On raw speed, Artificial Analysis measured the model at 91.3 tokens per second, ahead of the roughly 76 tokens per second median for comparable models, and actually faster than the 80 tokens per second SpaceXAI claims in its own materials.
That combination changes the calculation for teams running high volumes of agentic sessions. The Decoder's coverage framed it plainly: Grok 4.5 is priced so far below Fable 5 and GPT-5.5 that the benchmark gaps may not matter much for cost-sensitive workloads. Artificial Analysis's own comparison found Grok 4.5 costs roughly a fifth of what Claude Sonnet 5 costs per task on the Intelligence Index while scoring higher, a comparison LDS explored in more detail in its look at Sonnet 5's own cost-performance tradeoff against Opus 4.8.
Some practitioners are already voting with their workloads. Cursor CEO Michael Truell said the model had become the daily driver for many engineers on his own team, according to TechTimes' reporting, a notable signal from the company positioned to know its limitations best.
The unresolved question is not technical. It is institutional. SpaceXAI is separately facing litigation over harms linked to Grok's image generation tools, including a case LDS covered involving Tennessee teenagers suing xAI over Grok-generated deepfakes. None of that is a technical benchmark. But for enterprise buyers weighing whether to trust a vendor's own claims about its models, it is part of the same pattern this story documents: verify independently, because the launch materials are not the whole picture.
The Bottom Line
Strip away the framing and the facts are straightforward. Grok 4.5 is a genuinely capable, genuinely inexpensive model that ranks fourth among frontier systems, not first. Its launch-week disclosures include a doubled hallucination rate and a benchmark its own coding partner had to withdraw because the model had already seen the material it was being tested on.
Neither fact makes Grok 4.5 unusable. Together, they make it a model that requires the kind of independent verification Musk's launch post did not invite.
Two numbers defined Grok 4.5's first week. Both are 54. One is its score on the Artificial Analysis Intelligence Index, good enough for fourth place among frontier models. The other is the percentage of the time Artificial Analysis caught the model stating a wrong answer with confidence. SpaceXAI's announcement led with the first number. It never mentioned the second.
Sources
- Introducing Grok 4.5 — xAI/SpaceXAI Official Blog (Jul 8, 2026)
- Introducing Grok 4.5 — Cursor Official Blog (Jul 8, 2026)
- Scoop: Musk's SpaceXAI releases new model, Grok 4.5 — Axios (Jul 8, 2026)
- SpaceXAI releases Grok 4.5, which Elon describes as an 'Opus-class model' — TechCrunch (Jul 8, 2026)
- Grok 4.5 brings SpaceXAI to the intelligence frontier — Artificial Analysis (Jul 8, 2026)
- Grok 4.5 (high): Intelligence, Performance & Price Analysis — Artificial Analysis (accessed Jul 10, 2026)
- Grok 4.5 Cuts Coding-Agent Cost 80%: Near-Frontier Speed, Higher Hallucinations — Tech Times (Jul 9, 2026)
- Grok 4.5 is so cheap compared to Fable 5 and GPT 5.5 that benchmark gaps may not matter much — The Decoder (Jul 9, 2026)
- Grok 4.5 Launch: The Accuracy Trade-Off Nobody Headlined — Roo's Newsletter (Jul 9, 2026)
- SpaceX bets $60 billion on Cursor to catch OpenAI and Anthropic — The Decoder (Jun 16, 2026)
- Grok 4.5 Tops Agent Test, Backing Musk's Opus-Class Claim — BeInCrypto via Yahoo Tech (Jul 9, 2026)
- Grok 4.5 Beats Fable 5 And Opus 4.8 In Agent AI Test With 51.4% Score — Yellow.com (Jul 9, 2026)