Grok 4.5 'beats Opus' on exactly zero benchmarks

Here is the entire evidentiary basis for 'Grok 4.5 rivals Claude Opus': one post on X, from the man who owns the model, citing engineers who work for the man who owns the model. That is the story. Everything else is scale and vibes.

On 28 June 2026, Elon Musk announced that Grok 4.5 had entered private beta at SpaceX and Tesla and said early evaluations put it close to, or above, Anthropic's Claude Opus. Notice what is missing. No benchmark numbers. No system card. No independent evaluation. The 'evaluators' are SpaceX and Tesla engineers — employees of two companies Musk controls — so the test is graded by the home team, in a stadium the home team owns.

The one hard number on offer is size. Grok 4.5 runs on V9, xAI's ninth-generation architecture, which finished pre-training on 26 May 2026 at 1.5 trillion parameters — about three times the v8-small model serving Grok on X today. xAI reportedly added Cursor coding data to steer it towards code. Bigger and code-tuned is a reasonable bet for a better model. It is not proof of one, and a parameter count has never won a single head-to-head. Plenty of large models have lost to smaller, better-trained ones.

Rivals ship system cards for a reason: they document how a model was tested and where it breaks. xAI shipped a claim instead. And the timing is convenient — the beta lands as xAI promises a new foundation model every month through end-2026, with V9 variants feeding a larger Grok 5 on Colossus 2. A monthly hype cadence needs a monthly headline, and 'close to Opus' is a good one whether or not it turns out to be true.

None of this means Grok 4.5 is bad. It might be excellent. The point is narrower and harder to dodge: right now there is no way for anyone outside Musk's companies to know, and xAI has chosen to keep it that way. Treat the Opus comparison as a claim under dispute, not a fact — and wait for a leaderboard before you repeat it.

Zoom out and the pattern is the actual product. xAI has turned model releases into a content strategy: a monthly drop, a superlative claim, no system card, and a benchmark table that never quite materialises before the next drop buries the last. It's a treadmill designed to keep Grok in the conversation with OpenAI and Anthropic without ever having to withstand the scrutiny those labs invite by publishing. And it works — we're writing about it, you're reading about it — which is precisely why it should be graded on a curve. The second thing worth naming: using SpaceX and Tesla as the evaluation ground isn't just convenient, it's a moat and a conflict at once. Musk gets real production workloads to tune on, which is a genuine advantage, but he also gets to grade the exam, pick the questions and withhold the marksheet. A capable model tested only inside the founder's own companies is a capability claim with the verification deliberately removed. None of that makes Grok 4.5 bad. It makes it unfalsifiable — and an unfalsifiable frontier claim is worth exactly what you can independently check, which today is nothing.