Musk Shouted "Opus-Class." The Independent Scorecard Quietly Said Fourth.

On July 8, 2026, xAI shipped Grok 4.5, a model built specifically for coding agents. Elon Musk pinned a confident label on it right at launch: "Opus-class." That confidence didn't survive the day. On the Artificial Analysis Intelligence Index — the standardized, independent scorecard everyone in the field now watches — Grok 4.5 landed at fourth place with a score of 54, sitting behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. The model Musk called "Opus-class" was, in fact, ranked below Opus itself.

If that were the whole story, it'd be a shrug and a "the marketing ran a little hot." But it wasn't. The hits kept landing. The hallucination rate had doubled, from 25% on the prior Grok 4.3 to 54%. Cursor came out right after launch and admitted that a snapshot of its own codebase had accidentally ended up in Grok 4.5's training data. And the old complaint that Grok gives the most extreme answers of any tested model on political questions flared back up. Benchmark gap, hallucination spike, data contamination, political bias — four separate cracks in credibility, all opening at once.

Here's the twist, though: Grok 4.5 is not a bad model. It's dramatically cheaper than its rivals, and on one key capability it tops the entire board. So this is a story about a duality — a model that's genuinely good and genuinely hard to trust at the same time. It was a week where developers cheering a fast, cheap frontier model ran head-on into verifiers asking whether any of the numbers could be believed.

xAI, and Its New Face Called "SpaceXAI"

xAI is the AI company Musk founded in 2023. Armed with X's (formerly Twitter's) real-time data firehose and the massive "Colossus" GPU cluster, it barged late into the frontier race that OpenAI and Anthropic had staked out first. One thing worth noting in this launch: a lot of outlets started calling the company "SpaceXAI." As xAI's infrastructure and capital get more tangled with the SpaceX orbit inside Musk's empire, the name has shifted — and Grok 4.5 was introduced as the first product to come out of that combined resource base.

Anyone who knows Musk's playbook will recognize the "Opus-class" move. He always packages his products in superlatives. It happened with Grok 3, it happened with Grok 4 — "world's best" is never far behind. The problem is that the AI industry now has independent shops that fact-check that claim in real time. Outfits like Artificial Analysis don't run the cherry-picked benchmarks a company reads off its own slide; they push a standardized evaluation suite through every model and rank them. The gap between Musk's marketing and that cold scorecard is exactly where this controversy started.

For xAI, Grok 4.5 is also a signal of a pivot. Where the old Grok traded on being the "no-guardrails rebel chatbot," this release aimed squarely at a practical market: coding agents. Automated code generation, tool calls, multi-step task execution — the stuff that pulls real money out of developer wallets. A company that used to live on being provocative has now stepped onto a stage where it has to prove itself on results.

A Low Price, a Top Score, and the Mines Buried Underneath

Start with the price tag and you'll see why developers perked up. $2 per million input tokens, $6 per million output tokens. That's conspicuously cheap for a frontier-tier model. xAI claimed this pricing could cut coding-agent operating costs by up to 80%, and in benchmark runs it handled tasks at roughly $2.49 each — a figure that got compared to Claude Code's $11.80 and earned Grok 4.5 the nickname of the "value coder."

And the performance isn't uniformly weak, either. It's fourth on overall intelligence, sure — but on agentic tool use, Grok 4.5 took the number-one spot on the entire Artificial Analysis board. Deciding when and how to invoke a tool, then walking through multiple steps on its own — that's the single most important capability for a coding agent, and Grok 4.5 scored highest on it. Cheap and first in agentic ability? For a working engineer, there's no reason not to at least try it.

The trouble is the numbers buried under that shiny front. Accuracy climbed from the prior generation's 35% to 52% — but the hallucination rate leapt from 25% to 54%, more than doubling. This is the classic "the more it knows, the more confidently it's wrong" pattern. As the model's knowledge grows, so does its tendency to fabricate confidently about the things it doesn't know. In a domain like coding — where a plausible-but-wrong answer is more dangerous than an obviously wrong one — that 54% weighs a lot heavier than the price tag.

Metric Grok 4.3 Grok 4.5 Note
AA Intelligence Index rank — 4th (score 54) Behind Fable 5, GPT-5.5, Opus 4.8
Agentic tool use — #1 on the whole board Core coding-agent skill
Accuracy 35% 52% Improved
Hallucination rate 25% 54% More than doubled
Input price (per 1M tokens) — $2 Cheap for frontier tier
Output price (per 1M tokens) — $6 Cheap for frontier tier
Cost per task (benchmark) — ~$2.49 vs. Claude Code $11.80

And one more. Grok 4.5's whole sales pitch was that it trained on "trillions of tokens of Cursor coding data" — yet the benchmark Cursor itself built had to be excluded from the results because of contamination. The home-field evaluation it should have been proudest of got nullified by its own hand. Which brings us to the next chapter.

Who Won, and Who Took the Hit

The clearest winners are cost-sensitive developers and startups. At $2.49 a task, a team running thousands of agent calls a day sees its monthly bill shift by an order of magnitude. The number-one ranking on agentic tool use is especially attractive for people running automation pipelines that "write code, run tests, and fix it" — not just chat. On price-to-performance alone, Grok 4.5 genuinely moved the board.

Cursor sits in an awkward spot. On one hand, it slotted a model trained on its own coding data straight into its own editor and got to position it as "the model best tuned to our tool" — which is why Cursor's CEO called it the team's "daily driver." On the other hand, by disclosing the training-data contamination itself, Cursor undercut the credibility of its own benchmark. It banked the transparency points but also caught the backlash of "so what were all those CursorBench scores worth, then?"

The one who took the biggest hit, ironically, is Musk's own trust capital. Without that "Opus-class" line, a fourth-place finish would have read kindly as "a cheap fourth-place model, actually pretty usable." But by nailing down the superlative, he turned the independent fourth-place ranking into instant evidence of overselling. He raised the bar himself and then came in under it. Stack the hallucination, contamination, and bias controversies on top, and the individual issues blur into a bigger question: how much should anyone trust a number xAI publishes?

Quietly smiling, meanwhile, are Anthropic and OpenAI. Musk personally summoned Opus and GPT-5.5 as his points of comparison — and both of them sat above Grok 4.5 on the index. His rival did the "we're still on top" advertising for them, free of charge.

We've Seen This "Gap" Before — the Wins and the Faceplants

The gap between benchmark marketing and independent verification is an old song in AI. You don't have to look far. Startup after startup has unveiled a model with a homemade "we beat GPT-4" benchmark, only to slide down the rankings the moment a standardized external evaluation was applied. First place on the problem set you picked, mid-pack on the problem set someone else picked — that trap is common knowledge among industry insiders by now.

Data contamination isn't new, either. Several past models have been accused of having "test problems leak into the training data (contamination) and inflate the benchmark scores." In most cases the company denied it to the end or waffled. Against that backdrop, Cursor's proactive disclosure is unusually honest. Coming out first with "our code got into the training set and gave it an edge on CursorBench, so we're excluding this score" is a rare level of transparency by industry standards. The catch is that honesty isn't the same as restored trust.

The political-bias controversy is closer to a chronic condition for Grok. Earlier Grok generations were repeatedly flagged for an unusually high rate of extreme, one-sided answers to political questions. One analysis found Grok returning "strongly left or strongly right" extremes on a large share of political prompts; another found roughly 40% of its responses carrying only left-leaning arguments and 33% only right-leaning ones — a whipsawing distribution. Layer on the recurring circumstantial claims that Musk has personally intervened in Grok's outputs, and the suspicion of "did he tweak it again this time?" attaches itself automatically.

The lesson from all that history is a single one: once you get branded as "hard to trust the numbers," everything you ship afterward comes with an asterisk, no matter how good it is. Grok 4.5 is standing on exactly that fork in the road right now.

How the Competition Counters

Anthropic doesn't need to react loudly. Musk summoned Opus himself, and the index showed Opus 4.8 sitting above Grok. Anthropic actually gets a nice contrast effect out of it — "we let the rankings do the talking." That said, since Grok 4.5's weapon is clearly price, over the medium term Anthropic may well defend with a lower-cost coding tier or an agent-specific plan. Lean too hard on a performance edge and you risk losing the day-to-day volume work to a model that's "a bit weaker but five times cheaper."

OpenAI can quietly enjoy GPT-5.5 sitting ahead on the index, but ceding the number-one spot on agentic tool use to Grok has to sting. The coding-agent market is one of the most lucrative battlegrounds for the next several years, so expect OpenAI to push its Codex line toward nailing tool-calling ability and cost efficiency at the same time.

The most interesting move is Cursor's next one. With its own benchmark nullified by contamination, Cursor has no choice but to ship the "major CursorBench update" it already teased. Whatever score Grok 4.5 pulls on that fresh, uncontaminated evaluation will be the real ending to this whole affair. Do well on a clean board and it gets rehabilitated as "yeah, it really is a good model." Drop sharply and it's confirmed as "yeah, it was the contamination."

Then there's the regulatory variable. Grok 4.5 is classified under the EU AI Act as a general-purpose AI model with systemic risk, meaning xAI has to complete model evaluations, adversarial testing, incident-reporting procedures, and cybersecurity assessments before it can ship. With GDPR piled on top, the EU launch — originally slated for July 8–9 — got pushed to mid-July. How smoothly xAI clears the same regulatory gate its rivals already walked through will help decide the European fight.

So What Actually Changes

For developers, there's one more option on the menu — with conditions. A cheap model with strong agent chops is a clear plus. But a 54% hallucination rate is also a warning: "don't take the output at face value; always wrap it in a verification loop." Use it for its value on prototyping and high-volume repetitive work, but for logic or security-critical code where factual accuracy is life-or-death, stack two or three layers of verification. "Cheap" and "trustworthy" are different axes — keep that straight.

For investors and industry watchers, this episode is a textbook case of "benchmark inflation." It's a vivid demonstration of why the habit of placing a company's announced numbers next to the independent evaluation numbers matters. When you see a model claim first place on its own benchmark, checking whether that holds up on a standard index — and whether there's any whiff of data contamination — is now table stakes.

For general users, the political-bias part is what lands hardest. If you use Grok for information lookup or summarizing, filter its answers on political and social issues especially carefully. This is a model that's been repeatedly flagged for a high rate of extreme, one-sided answers. With the added suspicion that its owner has intervened in outputs, there's a bigger chance than with other models that a question you asked expecting "neutral facts" comes back skewed. Know that going in.

Bottom line: Grok 4.5 is a "the performance is real, the trust is a question mark" model. Its concrete strengths — price and agent ability — are undeniable, but four question marks now sit right on top of them: hallucination, contamination, bias, and hype. The next few weeks, especially Cursor's new benchmark and the EU launch, will decide whether those question marks get erased or set in stone.

🥄 Three Things You're Probably Wondering

— So can I actually use Grok 4.5 for my coding? You can, but bolt on a verification loop, no exceptions. It's cheap and it's number one on the board for agent ability, which makes it attractive for prototyping and repetitive work. But with the hallucination rate up at 54%, shipping its output verbatim is risky. Treat it as a "fast draft generator" that a human filters once more — that's the realistic fit for now.

— Was the "Opus-class" thing just a lie? Not quite a lie — closer to an exaggeration. On specific metrics like agentic tool use, it genuinely scored highest. But on the composite intelligence index it's fourth, below Opus 4.8, so the "Opus-class across the board" vibe clashes with the facts. Partly true, overall inflated — classic marketing phrasing.

— How serious is the Cursor data contamination? It's a real hit to score credibility, but it doesn't mean the model itself is junk. What got contaminated is one specific evaluation, CursorBench, and that was pulled from the results entirely. The problem is the symbolism of "the home benchmark you should be proudest of got nullified." Whatever score comes out of the new benchmark Cursor promised will be the real verdict.

Further Reading

Figures are as of announcement and may change.