Grok 4.5 in One Line: Opus-Class Coding at Discount Tokens
Grok 4.5 is a large language model built by SpaceXAI and Cursor for coding, AI agent development, and long-context knowledge work, combining Opus-class benchmark performance with lower token-based pricing and deep integration into the Cursor editor to compete directly with other frontier coding assistants.
This launch is not just another model drop; it is a strategic swing at the economics of serious coding assistance. SpaceXAI, formed after SpaceX absorbed xAI, released Grok 4.5 as its first joint model with Cursor and its first major launch since going public. The model landed on 8 July as a frontier system tuned for coding, agentic workflows, and office knowledge work. At USD 2 (approx. RM9.2) per million input tokens and USD 6 (approx. RM27.6) per million output tokens, Grok 4.5 undercuts several rivals while claiming Opus-class intelligence. That combination, plus Cursor AI integration, makes Grok 4.5 less a chatbot and more an opinionated bid to rewrite how developers pay for and use high-end coding assistants.
Benchmarks: Near the Frontier, With an Efficiency Twist
On coding assistant benchmarks, Grok 4.5 plays in the same league as the current giants but does not dominate every chart. SpaceXAI positions it as “an Opus-class model, but faster, more token-efficient and lower cost,” competing more with Anthropic’s Opus 4.8 generation than the very latest Claude releases. Benchmark data shows Grok 4.5 edging Opus 4.8 on Terminal-Bench 2.1 and beating both Opus 4.8 and GPT‑5.5 on DeepSWE 1.0, while losing ground on SWE-Bench Multilingual and SWE-Bench Pro and trailing Fable 5 overall. Cursor reports that on DeepSWE and SWE-Bench Pro, Grok 4.5 matches or outperforms top rivals while emitting fewer output tokens per task, improving practical speed and cost. In other words, its proposition is not “best IQ score” but “good enough IQ, fewer retries, shorter answers, and lower bills.”
The 500,000-token context window gives Grok 4.5 headroom for long pull requests, multi-file refactors, and sprawling agent sessions. That capacity will tempt teams running Terminal-Bench-like workflows in production, but it also raises the usual question: does a giant context actually pay off, or does it invite expensive, unfocused prompts? SpaceXAI acknowledges that requests above 200,000 tokens can face higher context pricing, which makes disciplined prompt design and tooling critical. Still, for developers juggling large monorepos and long-running agents, Grok 4.5’s benchmark profile plus massive context gives it strong “default model” potential—especially when paired with Cursor’s editor ergonomics.
Pricing: A Calculated Strike in the AI Model Pricing Comparison
Grok 4.5’s real weapon is its token economics. Official docs list USD 2 (approx. RM9.2) per one million input tokens, USD 0.50 (approx. RM2.3) per one million cached input tokens, and USD 6 (approx. RM27.6) per one million output tokens. Cursor also shows a faster variant at USD 4 (approx. RM18.4) input and USD 18 (approx. RM82.8) output. In the current AI model pricing comparison, that places Grok 4.5 below Anthropic’s Claude Opus 4.8 at USD 5 (approx. RM23) per million input and USD 25 (approx. RM115) per million output, while roughly matching OpenAI’s GPT‑5.6 Luna on output tokens at USD 6 (approx. RM27.6) but not on the USD 1 (approx. RM4.6) input rate. As one quotable summary: “Grok 4.5 is priced at USD 2 per million input tokens and USD 6 per million output tokens, well below rival flagships.”
Of course, raw token prices do not tell the whole story. SpaceXAI’s own docs stress that tool calls, web and X search, code execution, and long contexts can add separate fees beyond tokens. They frame the relevant metric as cost per completed task: finishing a pull request, spreadsheet model, research job, or document workflow after retries and tool invocations. Grok 4.5 is not the cheapest model in every stack; SpaceXAI notes that lower-tier and open-weight models—and even Grok 4.3—can have lower per-token rates. What SpaceXAI is really asserting is that for hard, multi-step coding and office tasks, Grok 4.5 hits a sweet spot: high enough quality to avoid endless retries, cheap enough to roll out across a team without CFO panic.
Cursor AI Integration and Agentic Workflows: Built Where Developers Live
Grok 4.5’s tight Cursor AI integration is where the launch feels most strategically dangerous for rivals. Cursor has rolled out Grok 4.5 as a core engine inside its coding environment, so developers can use the model directly in their editor on desktop, web, iOS, CLI, and SDK without juggling separate tools. The model is also accessible through Grok Build and the SpaceXAI console using an API key. For AI agent development, this means end-to-end flows—code search, refactors, tests, tool calls—can run in one place, with Grok 4.5 handling long-running agent sessions that read large codebases and iterate for minutes at a time. SpaceXAI’s docs describe Grok 4.5 explicitly as a model for coding, agentic tasks, and knowledge work, and Cursor’s focus on real developer workflows turns that description into concrete product decisions.
The training pipeline reinforces this focus. SpaceXAI and Cursor jointly trained Grok 4.5 as a mixture-of-experts model using trillions of tokens from real Cursor user sessions, capturing how engineers write, review, and debug code and how coding agents move through codebases in practice. Training ran across tens of thousands of Nvidia GB300 GPUs, with aggressive data filtering, deduplication, and domain-focused curation. This is a clear bet that usage data from a coding editor matters more than static code dumps. For developers, the implication is simple: Grok 4.5 should feel tuned to real workflows—incremental edits, code review comments, and half-broken branches—rather than academic puzzle-solving. Whether that advantage holds over time will depend on how quickly rivals match this tight IDE-model feedback loop.
Competition With GPT‑5.6 and the Road Ahead for Developers
Timing matters. Grok 4.5 launched in the same week that OpenAI began rolling out GPT‑5.6 more broadly, creating a crowded window for frontier releases and intensifying the fight over developers’ coding workflows. SpaceXAI is not claiming absolute supremacy in raw intelligence; instead, it is making an explicit economic argument built on speed, token efficiency, and cost. Meanwhile, the model is not yet available in the EU, with API and product access expected to open there in mid‑July, which temporarily blunts its global reach. Still, availability across Grok Build, Cursor, and the SpaceXAI console gives non‑EU teams a fast path to production.
This launch is also the first major output of SpaceXAI’s roughly USD 60 billion (approx. RM276 billion) acquisition of Anysphere, Cursor’s parent company, and follows Elon Musk’s overhaul of xAI’s leadership and pivot to acquiring Cursor. It signals a clear strategy: own both the model and the primary coding surface, then compete on developer-first pricing. For ordinary users, the upside is tangible. Teams running agentic workloads that read large codebases, call tools, and iterate for minutes at a time may see per-token savings compound across their development organization, especially if Grok 4.5 lowers the cost of finishing pull requests, spreadsheet models, research tasks, and office document workflows when retries and long prompts are factored in. The frontier race has shifted: the winner is no longer the model with the tallest benchmark bar, but the one that closes tickets fastest without blowing the cloud bill.
