Grok 4.5: A Coding Model Built to Win on Price, Not Bragging Rights
Grok 4.5 is a large language model from SpaceXAI designed specifically for coding and agentic workloads, pairing frontier‑class capability with sharply lower token prices to give developers a cost‑effective alternative to premium Opus and GPT models for sustained software engineering tasks. Instead of chasing the highest possible intelligence score, SpaceXAI is betting that most engineering teams care more about cost per completed task than leaderboard bragging rights. With pricing of USD 2 (approx. RM9.20) per million input tokens and USD 6 (approx. RM27.60) per million output tokens, Grok 4.5 deliberately undercuts flagship rivals while still claiming “Opus‑class” performance on coding benchmarks. In other words, it is engineered less as a conversational assistant and more as a workhorse AI coding model that aims to reshape AI coding pricing for teams that run agents all day long.
Cursor Integration Turns Training Data Into a Developer Product
The most interesting part of Grok 4.5 is not just its size; it is where the model learned to code. This release is the first major product of SpaceXAI’s roughly USD 60 billion (approx. RM276 billion) acquisition of Anysphere, the company behind the Cursor IDE, and Cursor’s interaction data played a central role in training. SpaceXAI trained Grok 4.5 across tens of thousands of Nvidia GB300 GPUs with careful filtering, deduplication, and domain‑focused curation, using real developer behavior—how engineers write, review, and debug code—to shape the model. Now that same model is embedded as a core engine inside Cursor’s AI coding environment, giving developers access directly in their editor rather than through yet another separate tool. This is what makes Grok 4.5 more than an abstract Opus alternative: the training pipeline and the Cursor integration form a feedback loop for AI agent development, where real usage improves the model and the model, in turn, powers richer coding agents.
Opus-Class Benchmarks at a Quarter of the Output Cost
SpaceXAI is blunt about its ambition: Elon Musk calls Grok 4.5 “an Opus‑class model, but faster, more token‑efficient and lower cost.” The numbers give that claim teeth. Grok 4.5 runs on the new V9 foundation with 1.5 trillion parameters, about three times larger than the v8‑small system that powered Grok 4.3. It supports up to 500,000 tokens of context, enabling long‑horizon coding sessions over large repositories. On coding tests such as DeepSWE and SWE‑Bench Pro, Grok 4.5 matches or outperforms top rivals while generating fewer output tokens per task, which directly improves cost efficiency for iterative work. According to one benchmark, it reached 64.7% on SWE‑Bench Pro and handled automated coding tasks on CursorBench for about USD 1.51 (approx. RM6.95) per task, while charging only USD 6 (approx. RM27.60) per million output tokens compared with Claude Opus 4.8 at USD 25 (approx. RM115) per million output tokens.
AI Coding Pricing Becomes the Battlefield for Agentic Workloads
Grok 4.5’s USD 2 (approx. RM9.20) input and USD 6 (approx. RM27.60) output token pricing is not a minor tweak; it is a frontal attack on how AI coding is priced today. Claude Opus 4.8 sits at USD 5 (approx. RM23) for input and USD 25 (approx. RM115) for output per million tokens, while OpenAI’s priciest model reaches USD 5 (approx. RM23) and USD 30 (approx. RM138) and its cheaper option lands at USD 1 (approx. RM4.60) and USD 6 (approx. RM27.60). For agentic workloads that read entire codebases, call tools, and iterate for minutes, that output gap compounds fast across a team. One benchmarking firm already ranked Grok 4.5 fourth on a real‑world agentic knowledge‑work index while measuring its cost per completed task at a fraction of the leaders. The message is clear: if you are paying frontier prices for coding agents, Grok 4.5 forces you to justify the premium on reliability and outcome quality, not on raw token counts.
What Cheaper Frontier Coding Means for Developers and Users
Grok 4.5 is available through the SpaceXAI console via API key, through the Grok Build agent, and inside Cursor, with EU access expected later in July. Engineers can also use it across Cursor plans, command‑line workflows, and other developer platforms, with a clear focus on coding and agentic tasks rather than casual chat. SpaceXAI claims the model is not yet using proprietary software that could make it run twice as fast or better on Nvidia’s latest GB300 chips, which hints that the pricing pressure may increase further once that optimization layer is in place. For businesses, this points toward cheaper AI agents that can work at scale and puts direct pressure on rivals to explain their higher prices. For consumers, the impact is indirect but real: when automated engineering becomes cheaper, products can be built and updated faster and may cost less to run. If Grok 4.5 holds up in messy real repositories, AI agent development shifts from "most intelligent" to "best value per shipped feature."
