Grok 4.5: Musk launches "Opus-class" model at a fraction of the price

What it's about

SpaceXAI, the company formerly known as xAI and now publicly listed, has released Grok 4.5. According to TechCrunch, it is the first model since the IPO a few weeks ago. Elon Musk announced it on X on Wednesday, July 8, 2026, saying Grok 4.5 would be publicly available the following day. His wording: it is "an Opus-class model, but faster, more token-efficient and cheaper." The model had previously run in a private beta program, and the stable version until then was Grok 4.3. With "Opus," Musk is alluding to Anthropic's Claude Opus.

What the model is supposed to do

In a blog post, SpaceXAI describes Grok 4.5 as a workhorse for typical automation tasks, according to TechCrunch: programming and app building, office and admin work, research, writing and other knowledge work. The company touts "twice the token efficiency" compared with other leading models. That would matter, since token costs are becoming a bigger issue for customers. According to Tech Insider, the developer notes name four focus areas: coding (fine-tuned for Rust and C/C++), multi-step agentic tasks, office work such as multi-sheet Excel models with web research, and research and writing through to PowerPoint and Word.

Technically, Grok 4.5 is based on a new foundation model called V9 with 1.5 trillion parameters, according to Tech Insider. That is roughly three times larger than the earlier "v8-small" architecture (a figure from CoinGape; the size is confirmed by Writingmate). Training reportedly ran on tens of thousands of Nvidia GB300 GPUs in the Colossus cluster in Memphis. According to the vendor's documentation, the context window is 500,000 tokens. The data mix is new: the model was trained together with developer data from Cursor, the coding tool that, according to Tech Insider, was folded into the company. It is the default model in "Grok Build," SpaceXAI's development platform, and according to Cursor is available on all plans.

Price comparison

According to Tech Insider, a monthly volume of 50M input and 10M output tokens costs about $160 with Grok 4.5, $500 with Opus 4.7 and $550 with GPT-5.6 Sol. OpenAI's mid-tier "Terra" has no known pricing yet.

Benchmarks and first impressions

SpaceXAI published benchmark results that, according to TechCrunch, show Grok is competitive but just below the top. Musk later said: "Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but significantly faster." He added on X that Grok 4.5 doesn't yet use the company's in-house C/C++ inference software for GB300 hardware, and that a doubling of speed or more is probably possible.

NextBigFuture collects voices from X (individual opinions, not studies). One developer writes that he uses Grok 4.5 as the default for all coding work, and that costs could fall by 50–80 percent. According to BridgeMind, Grok 4.5 ranks third on CursorBench at 66.7 percent, behind Fable 5 Max at 70.5 percent, at $1.51 per task instead of $17.32. Investor Gavin Baker calls it "Pareto-dominant" for coding based on the numbers, with the "vibes" still to come. NextBigFuture's Brian Wang speculates that Cursor-SpaceXAI revenue could rise sharply. That is an unsupported forecast.

Market context

The release lands in a packed week. OpenAI is releasing GPT-5.6 publicly after, according to Mashable and TechCrunch, its launch was earlier limited by the US government. According to Mashable, Anthropic's strongest model, Fable 5, was also temporarily taken offline after a US government intervention and is available again. Competitors' prices are an argument for Grok. However, xAI's positioning and the price comparison both refer to Opus 4.7, while Anthropic already offers Opus 4.8, according to Tech Insider. Mashable lists Opus 4.6 as the current version, so the sources contradict each other here.

Our take

Grok 4.5 competes mainly on price. If it can deliver near-Opus quality at a third to a quarter of the cost, it would interest developers with large token volumes. So far, that is backed only by vendor claims, Musk's internal assessment and early user reports. According to Tech Insider, independent benchmarks against Opus 4.8 are missing. It is also unclear whether the lower token count per task ("double efficiency") holds up in everyday use. Grok 4.5 is not an open-weights model. It is available only through the service and API. And because SpaceXAI is publicly listed, every launch carries extra financial weight.