Grok 4.5 Is Here: How SpaceXAI Is Undercutting Claude and GPT on Price Without Giving Up on Performance

Elon Musk's AI operation just made its first major release since folding xAI into SpaceX and the story isn't about who built the biggest model. It's about who built the cheapest one that still holds its own.

On July 8, 2026, SpaceXAI launched Grok 4.5, calling it the company's most intelligent model to date and its strongest built specifically for coding, agentic tasks and knowledge work. It's also notable for being the first model SpaceXAI built jointly with Cursor, the AI coding platform SpaceX agreed to acquire for a reported $60 billion in an all-stock deal.

Built On Real Developer Data, Not Just Code Repositories

Where Grok 4.5 diverges from a typical model release is in how it was trained. Rather than leaning purely on static code repositories, SpaceXAI and Cursor trained the model on trillions of tokens of actual developer-agent interaction data: real bugs, real debugging sessions, and real architecture decisions rather than just finished code. The idea is that a model learns more from watching how engineers actually work through a problem than from studying only the polished result.

Training itself ran across tens of thousands of Nvidia GB300 GPUs, with heavy investment in data filtering, deduplication and quality scoring. Reinforcement learning played a major role too, with Cursor's engineers designing problems specifically difficult enough to keep challenging a frontier-level model, since tasks that no longer trip up a model stop teaching it anything new.

The Real Headline: Price

The number most engineering teams will care about is cost. Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens. For comparison, Anthropic's Claude Opus 4.8 runs $5 and $25 for the same, and Claude Fable 5 runs considerably higher still, while OpenAI's GPT-5.6 Luna sits closer to Grok on price. Musk has directly compared the model to Anthropic's flagship, describing it as an Opus-class model that is faster, more token-efficient and lower-cost.

The efficiency gains show up beyond sticker price too. Grok 4.5 reportedly uses roughly four times fewer output tokens than Opus 4.8 to complete equivalent tasks on SWE-Bench Pro, and independent efficiency comparisons show a similar pattern elsewhere. Since agentic workflows can loop through a model hundreds of times on a single task, fewer tokens per step compounds into meaningfully lower real-world cost, not just a better number on a pricing page.

Where It Lands on Benchmarks

Grok 4.5 doesn't top every chart, and SpaceXAI isn't really claiming it does. Independent benchmarking from Artificial Analysis placed it fourth on its Intelligence Index, behind Fable 5, GPT-5.5 and Opus 4.8. On coding-specific tests the picture is mixed: it leads on SWE Marathon and ranks first on Harvard's Legal Agent Benchmark, but trails Opus 4.8 and Fable 5 on tests like SWE-Bench Pro and DeepSWE.

The company's own pitch leans less on topping leaderboards and more on the ratio of capability to cost and speed. It's worth noting one transparency wrinkle flagged in coverage of the launch: an earlier snapshot of Cursor's own codebase was accidentally included in Grok 4.5's training data, which may have given it a benchmarking edge on tasks tied to that codebase specifically.

Speed and Where You Can Use It

Grok 4.5 runs at roughly 80 tokens per second, which SpaceXAI and outside reporting both describe as faster than comparable frontier models. Musk has also suggested that a custom inference stack still in development could double that speed once deployed.

The model went live broadly on day one: it's available through the SpaceXAI API under the model ID grok-4.5, inside Grok Build (SpaceXAI's own coding agent), and across all Cursor plans, with third-party access already live on platforms like OpenRouter, Vercel, Cloudflare, Snowflake and Databricks Mosaic. SpaceXAI is also positioning it as a broader knowledge-work tool, capable of building multi-sheet Excel models with integrated web research and producing polished PowerPoint and Word content, not just handling code.

One notable gap: Grok 4.5 isn't yet available to users in the EU, with SpaceXAI indicating that access is expected by mid-July.

The Bigger Picture

Grok 4.5 arrived just a day ahead of OpenAI's own GPT-5.6 launch, in a stretch of a few weeks that has seen several major labs ship frontier-class models in close succession. What makes this release distinct isn't a claim to the top spot. It's a bet that token efficiency and inference cost are quickly becoming just as important to engineering teams as raw benchmark scores, especially as agentic workflows scale usage into the hundreds or thousands of model calls per task.

For developers and platform teams evaluating AI coding tools, Grok 4.5 adds a genuine third option to the conversation, one where the pitch isn't "the smartest model available" but "smart enough, priced to actually run at scale."