Grok 4.5 lands: Opus-class claims, cheaper, unproven

SpaceXAI, the company formerly known as xAI, released Grok 4.5 within the last few hours, its first model since going public and acquiring the AI coding startup Cursor, and Elon Musk is pitching it as an Opus-class coder that is faster, more token-efficient, and far cheaper than Anthropic's flagship. The catch, and our thesis, is that Grok 4.5 shipped on price and vibes rather than proof: it has no independent SWE-bench Verified score and was not submitted to Artificial Analysis, LMArena, or any recognized third-party evaluator. When a frontier model launches this loudly with only vendor-run numbers, the honest read is promising but unproven.

• Grok 4.5 is priced at $2 per million input tokens and $6 per million output, undercutting Claude Opus 4.8 at $5/$25 and matching OpenAI's cheaper tier.
• SpaceXAI trained it alongside Cursor, its new acquisition, and it is available today in Grok Build, the SpaceXAI console, and Cursor on all plans, though not yet in the EU (mid-July expected).
• The vendor benchmarks are mixed: SpaceXAI claims Terminal-Bench 2.1 of 83.3% (ahead of Opus 4.8's 78.9%) but trails on SWE-bench Pro, 64.7% to Opus 4.8's 69.2%.
• There is no SWE-bench Verified number and no independent eval, and Cursor disclosed a benchmark contamination issue, so treat the coding claims as unconfirmed for now.

What did SpaceXAI actually launch?

Grok 4.5 is being sold as SpaceXAI's smartest model yet, built to excel at coding, agentic tasks, and knowledge work, and it is the company's first release since two big corporate moves: going public and buying Cursor, the popular AI coding editor. Musk described it on X as an Opus-class model that is faster, more token-efficient, and lower cost, positioning it as a coding tool for developers rather than a consumer chatbot. It is available immediately in Grok Build, the SpaceXAI developer console, and inside Cursor across every plan, with free usage for a limited time. EU access is held back until roughly mid-July, a familiar pattern for frontier launches navigating European AI rules.

RelatedOpenAI retracts SWE-Bench Pro after finding 30% broken

Why is the pricing the real headline?

Because on cost, Grok 4.5 is genuinely aggressive. At $2 per million input tokens and $6 per million output, it sits far below Claude Opus 4.8 at $5 and $25, and it lands in the same neighborhood as OpenAI's cheaper GPT tier at roughly $1 in and $6 out. If the quality claims even partly hold, that price gap is the story, because agentic coding burns enormous token volumes and output tokens dominate the bill. SpaceXAI also says the model is served at around 80 tokens per second with roughly double the token efficiency of leading rivals on the same tasks, which compounds the savings. For a team running long autonomous coding jobs, a four-times-cheaper output rate is not a rounding error, it is the difference between viable and not.

Why do the missing benchmarks matter so much?

This is the part most launch coverage glosses over, and it is the reason to keep the champagne corked. Grok 4.5 arrived with company-run numbers only: a claimed number-one finish on Harvey's Legal Agent Benchmark, an 83.3% on Terminal-Bench 2.1 that edges Opus 4.8, and a 64.7% on SWE-bench Pro that actually trails Opus 4.8's 69.2%. What is absent is telling. There is no SWE-bench Verified score, the metric most of the industry treats as the coding yardstick, and the model has not been handed to Artificial Analysis, LMArena, or any neutral evaluator. Worse, Cursor disclosed that an earlier snapshot of its own codebase had accidentally leaked into training, so it excluded CursorBench results as contaminated. None of that means Grok 4.5 is bad. It means the evidence we have is exactly the evidence a vendor controls, and history says vendor charts flatter their own model. Until an independent lab runs it, the coding crown stays with the models that have been measured in the open.

What does it mean for the market?

The signal for investors runs through two names, one private and one very public. SpaceXAI itself is private, but its most direct listed proxy is Tesla (NASDAQ: TSLA), which holds a stake in the AI venture and whose bull case increasingly leans on Musk's AI ambitions; a credible, cheap frontier coder strengthens that narrative even before it is independently verified. The sharper read is on the pricing war it reopens: by putting an alleged Opus-class model at a fraction of Opus pricing, SpaceXAI pressures the margins of every API-first lab, which matters for how much cloud and inference spend flows to Nvidia (NASDAQ: NVDA), the common supplier underneath all of them. Notably, Grok 4.5 was trained on tens of thousands of Nvidia GB300 GPUs using the same compute SpaceXAI leases to competitors including Anthropic and Google, so xAI is now both a customer and a landlord in the AI arms race. The number to watch is not the launch-day chart, it is whether an independent SWE-bench Verified score lands near the vendor claims; that, not Musk's post, moves the competitive math.

Who is affected first?

Developers and coding-tool vendors feel it immediately. Cursor users get a cheap, capable model wired in on every plan, which is exactly why SpaceXAI bought Cursor: distribution. Rival coding assistants now face a price anchor that is hard to match without eating margin. Anthropic and OpenAI are pushed to justify premium pricing on proven quality, which is precisely where Grok 4.5's missing benchmarks cut against it. And EU developers are shut out for now, a reminder that regulatory timing still shapes who gets frontier tools first.

RelatedAnthropic's Claude Science targets neglected diseases

1. Jul 2025Grok 4 launches. xAI's first serious frontier push, strong on reasoning benchmarks.
2. Jun 2026Grok 4.5 enters private beta. Tested inside Tesla and SpaceX, no public access or independent benchmark.
3. Jul 8 2026Grok 4.5 ships publicly. First model after xAI goes public and acquires Cursor; Opus-class claim, $2/$6 pricing.
4. Mid-Jul 2026EU availability expected. Access widens; independent evals will decide if the claims hold.

What is next?

The next two weeks decide the narrative. If Artificial Analysis or a SWE-bench Verified run confirms Grok 4.5 near Opus-class at a quarter of the output price, this becomes one of the most disruptive launches of the year and forces an industry-wide price rethink. If independent numbers come in well below the vendor charts, the launch reads as marketing, and the Cursor contamination note will loom larger. Either way, expect SpaceXAI to lean on distribution through Cursor and free-tier access to build usage fast, because in a market where switching costs are low, being cheap and everywhere is its own strategy.

Our take

Grok 4.5 is the most interesting launch of the week and also the least proven, which is an unusual combination. The pricing is real and aggressive, the Cursor acquisition gives it distribution most challengers can only dream of, and an Opus-class model at a fraction of Opus cost would genuinely reshape the coding-tool market. But we do not review press releases, we review evidence, and right now the evidence is a set of charts the vendor drew, with the industry's standard benchmark conspicuously missing and a training-contamination footnote attached. That is why our leaderboard lists Grok 4.5 as verifying, not ranked: not a knock on the model, just an insistence that a number counts once someone other than its maker measures it. If the independent scores land where Musk says they will, we will move it up fast. Until then, the smart posture is to try it, because it is cheap, and to withhold the crown, because it has not been earned in the open.

Original analysis by GenZTech. Reporting informed by Axios.