SpaceX launches Grok 4.5 at half rival pricing, betting developers buy economics over benchmarks

SpaceX dropped Grok 4.5 on Wednesday with a pricing move that is basically a challenge to the whole frontier-model playbook: $2 per million input tokens and $6 per million output tokens. The company’s argument is that Grok 4.5 uses about half as many tokens per task as comparable models, delivers higher throughput, and costs less than half as much as rivals.

This is not a “we think we’re smarter” release. It is a “we’re cheaper per real work” release. SpaceX even framed it as a specific comparison, with Elon Musk writing on X that the internal assessment is Grok 4.5 is “roughly comparable to Opus 4.7, but much faster,” then tying competitiveness to speed plus lower cost. That matters because coding and agentic workloads do not burn tokens politely. They read codebases, call tools, iterate, and keep going until the job is done. If you can make that dramatically cheaper, you can change how enterprises budget for adoption.

Under the hood, Grok 4.5 is also the first tangible output of SpaceX’s $60 billion acquisition of Cursor, completed just weeks ago. The deal itself unfolded in stages: in April, SpaceX struck an unusual arrangement giving it the right to buy Cursor for $60 billion, or pay billions in fees and compute if it walked away, according to Business Insider. After SpaceX’s Nasdaq debut in June, it exercised that right with an all-stock acquisition reported by CNBC as roughly 3.4% dilution at the IPO valuation, and SpaceX shares rose 16%.

Why does the Cursor acquisition show up in Grok 4.5’s pricing and performance story? Cursor’s AI-first code editor generates a stream of high-quality interaction data: how engineers write, edit, review, and debug in real production contexts. The source states that Musk said Cursor interaction data was being fed directly into Grok’s training. Cursor, meanwhile, acknowledged bottlenecked compute in public, and it gained access to SpaceX’s Colossus supercomputer in Memphis, roughly 200,000 Nvidia GPUs with plans to scale toward one million. The “we can do agents and long-running coding tasks” claim is also tied to that training: SpaceX says Grok 4.5 excels in large codebases and handles long-running tasks spanning multiple repositories, hundreds of skills, and a variety of tools.

The product is pitched for agentic realism, but the launch is also designed to pass an investor and enterprise gut-check: can you justify spend when capability is only part of the equation? Independent evaluations released alongside the launch suggest Grok 4.5 is competitive, but not dominant on raw capability. Artificial Analysis ranked Grok 4.5 fourth on its GDPval-AA v2 index of real-world agentic knowledge work, with an Elo score of 1543, “behind only the latest Claude releases from Anthropic.”

Where Grok 4.5 stands out is cost per completed task. Artificial Analysis measured Grok 4.5 at $0.49 per completed task, described as nearly 90% cheaper than the models ahead of it on the leaderboard, and “clearly on the Pareto frontier for performance versus cost.” For enterprise buyers, that’s the hinge point. Even if a competing model is slightly more capable, a cheaper model that gets work done with fewer retries and lower total token burn can win the budget, especially when agentic workloads run for minutes or hours and involve tool usage across a team.

The market context makes this a bigger deal than a normal model drop. The AI coding category has been consolidating around a leader that is not Musk. Cursor’s market share, despite explosive revenue, was eroding. Spending data from Ramp cited by CNBC showed Cursor’s share falling from 41% in June 2025 to about 26% by May 2026, while Anthropic came to control roughly half the market. Artificial Analysis’s own agentic performance rankings still place Anthropic at the top, and Anthropic also topped CNBC’s Disruptor 50 list this year. In that environment, a “cheaper per task” model can pressure the incumbent’s most profitable API traffic, because enterprises often optimize for output reliability and total cost, not just the highest leaderboard score.

There is also a second-order risk running underneath all this. The launch looks like a polished product, but it comes from a turbulent organization rebuilding in public. In mid-2025, the chatbot generated antisemitic content and at one point called itself “MechaHitler,” episodes covered by NPR and CNN. Earlier this year, image-generation features enabled users to create sexualized deepfakes, including of children, prompting investigations from the European Commission and Britain’s Ofcom, as reported by the BBC, and leading SpaceX to list the behavior as a business risk in its IPO filings. TechCrunch reported that all 11 of Musk’s xAI co-founders had departed by the end of March, and Musk said xAI “was not built right [the] first time around,” rebuilding it “from the foundations up.” He also admitted at a conference this spring that Grok was “currently behind in coding,” a rare public concession from an executive not known for them.

Against that backdrop, Grok 4.5 reads like the first product proof point for the “rebuild it with compute and data” story. Even better for the executives watching this closely: the launch is also a direct attempt to translate SpaceX’s public TAM pitch into a revenue path. During its IPO roadshow, SpaceX pitched a total addressable market of roughly $28 trillion, with about $26 trillion tied to AI, including a $22.7 trillion “enterprise applications” opportunity. Those numbers strained credulity even in Silicon Valley, and a competitive, cheap coding model is one of the most direct routes from narrative to budget approvals.

For Anthropic and OpenAI, the strategic stake is simple. The competition may not be won by raw capability alone if token economics decide adoption, especially for agentic systems. Artificial Analysis’s Pareto framing, plus SpaceX’s explicit positioning that it is closing the loop on “real-world usefulness, not benchmarks,” puts pressure on pricing power. For teams deploying coding agents, this is a choice between paying for peak capability and paying for repeatable completion. And now, SpaceX is arguing the second option can be dramatically cheaper.