xAI drops Grok 4.5 for coding and agents with 80 TPS speed

While the rest of the frontier labs are fighting tooth and nail over fractional benchmark gains, SpaceXAI is playing a very different game: making frontier intelligence fast and dirt cheap.

SpaceXAI officially launched Grok 4.5, its new flagship model engineered specifically for software development, agentic workflows, and knowledge work. It is the company's first major model release since going public a few weeks ago.

Built alongside Cursor on tens of thousands of GB300s

The most notable detail about Grok 4.5 is that it was trained directly alongside Cursor. To build it, SpaceXAI deployed tens of thousands of NVIDIA GB300 GPUs, lean data curation, and an asynchronous reinforcement learning stack where agentic rollouts could run for hours across multi-step engineering tasks.

The real hook here is speed and token efficiency. Grok 4.5 clocks in at 80 tokens per second and reportedly solves tasks in under half the number of steps compared to rival flagships. According to data reported by Forward Future, Grok 4.5 averaged 15,954 output tokens on SWE Bench Pro tasks.

It handles difficult C/C++ and Rust problems, one-prompt full-stack apps using libraries like Three.js, and multi-sheet financial modeling in Excel complete with sticky notes. SpaceXAI also built plugins hooking it straight into Word, PowerPoint, and Outlook.

Elon Musk took to X to frame the launch, calling it an "Opus-class model" that their internal testing finds roughly comparable to Anthropic's Opus 4.7, only significantly faster and far cheaper. Pricing is set at $2 per million input tokens and $6 per million output tokens, with availability today in Cursor, Grok Build, and the SpaceXSI console, alongside a limited-time free tier. European Union access is slated for mid-July.

Why it matters

We think Grok 4.5 represents a clear shift in how frontier models will compete going forward: intelligence per second and intelligence per dollar.

For developers and enterprise teams, the math speaks for itself. Anthropic charges $5 per million input and $25 per million output tokens for Opus 4.7, while OpenAI's top-tier GPT-5.6 Sol costs $5 input and $30 output. At $2 and $6, combined with roughly double the token efficiency, running multi-turn autonomous coding agents suddenly becomes drastically cheaper. If you are already building inside Cursor, trying Grok 4.5 while it is free right now is a no-brainer.

SpaceXAI's go-to-market move here is remarkably clever. Instead of trying to claim every academic benchmark crown, they partnered directly with the editor where developers actually live. It turns Cursor into a massive distribution engine right as OpenAI rolls out GPT-5.6 Sol.

That said, there is an obvious catch. Benchmark data published alongside the launch shows Grok 4.5 still sitting just shy of absolute best-in-class performance. In complex software engineering, a blazing fast, cheap model that misinterprets an edge case still requires human triage, which eats up developer hours. We will soon find out whether teams prefer raw perfection or Grok's lightning-fast compromise.