SpaceXAI’s Grok 4.5 Undercuts OpenAI and Anthropic on Coding

SpaceXAI and Cursor released Grok 4.5 on Wednesday, the first jointly developed AI model to ship under Musk’s new SpaceXAI name and the first Cursor build trained on the Colossus supercomputer inside SpaceX. The model is live in Grok Build, inside Cursor on every plan, and on the SpaceXAI console, priced at $2 per million input tokens.

The launch lands weeks after SpaceX formally agreed to acquire Cursor’s parent Anysphere for $60 billion in stock, a deal expected to close in the third quarter of 2026. Internal coverage of the launch was carried by Thunder Tiger Europe Media on Wednesday. Neither company has called the model a test run, but the messaging is calibrated for cost rather than raw performance.

SpaceXAI and Cursor Ship Grok 4.5 as a Joint Model

SpaceXAI’s official launch page opens with the framing the company will lean on for the next quarter: “Today, we’re launching Grok 4.5, SpaceXAI’s smartest model built to excel at coding, agentic tasks, and knowledge work. It’s our strongest model ever and was trained alongside Cursor.” The site lists coding, engineering, math, and science as the training domains, and bills the release as SpaceXAI’s strongest model to date.

Behind the launch sits a rebranding that the two companies had telegraphed for months. XAI was formally folded into SpaceX in February 2026, and the account handle flipped to SpaceXAI on Monday, accompanied by a new logo. The release is the first major product to carry that name, and it ties the rebranded company directly to Cursor’s coding tool.

Distribution matters as much as the model. Grok 4.5 ships as the default in Grok Build, into Cursor on every plan, and through the SpaceXAI console for API access. The model is not yet available in the EU in any SpaceXAI products or the API console; EU availability is expected in mid-July, according to the launch page.

The Cost Pitch Is $2 and $6

The pricing is the headline. SpaceXAI is selling Grok 4.5 at $2 per million input tokens and $6 per million output tokens, a number designed to land against the two models Cursor was already selling against. The company ties the price to a token-efficiency claim that the launch page documents directly: Grok 4.5 resolves SWE Bench Pro tasks with 15,954 output tokens on average, about 4.2× fewer than Opus 4.8 max at 67,020.

The rivals are priced where their publishers state, with no rounding. Anthropic’s Opus 4.8 lists at $5 input and $25 output per million tokens; OpenAI’s GPT-5.6 Luna lists at $1 input and $6 output; Anthropic’s Claude Sonnet 5 is in an introductory window of $2 and $10 per million tokens through August 31, 2026, after which standard pricing is $3 and $15 per million (per Claude’s published pricing page). Grok 4.5 is the cheapest option that covers the entire coding range Cursor’s customers have been routed to.

Cursor as the Distribution Channel

Grok 4.5 has a built-in audience. Cursor said in November that it had crossed $1 billion in annualized revenue, and the coding tool was used by a roster of enterprise customers building software agents on top of the IDE. That user base is now pointed at Grok 4.5 by default.

Cursor’s own position had been weakening even before the SpaceXAI deal closed. CNBC reported that Cursor’s market share had declined from 41% in June 2025 to about 26% in May 2026, according to spending data from Ramp, with Anthropic now controlling roughly half of that category. The Grok 4.5 push is timed to slow that slide by giving Cursor a cheaper default to ship.

Excited to partner with the SpaceX team to scale up Composer. A meaningful step on our path to build the best place to code with AI.

That quote is from Cursor chief executive Michael Truell, posted to X at the time of the April framework deal and reproduced in the Cursor acquisition deal terms. CNBC reported that SpaceX formally agreed to acquire Anysphere, Cursor’s parent, for $60 billion worth of stock on June 16, 2026, and expects the merger to close during the third quarter of the year, subject to regulatory approvals. If the deal collapses, SpaceX has agreed to pay Cursor a $1.5 billion termination fee plus $8.5 billion in computing resources, per the company’s IPO filings.

Where Grok 4.5 Wins and Where It Trails

Elon Musk pitched the model directly against Anthropic’s top family in his X post on launch day. “It is an Opus-class model, but faster, more token-efficient and lower cost,” Musk wrote, attaching a chart that put Grok 4.5 ahead of Opus 4.8 on several benchmarks.

The numbers SpaceXAI itself published tell a more specific story. DeepSWE 1.0 (run by Datacurve with each provider’s harness): Fable max 66.1%, GPT-5.5 xhigh 64.31%, Grok 4.5 62.0%, Opus 4.8 max 55.75%, Opus 4.7 max 40.12%. SWE Marathon resolution rate at pass@1: Grok 4.5 29.0%, Opus 4.8 max 26.0%, Fable max 24.0%. Terminal Bench 2.1: Fable max 84.3%, GPT-5.5 xhigh 83.4%, Grok 4.5 83.3%, Opus 4.8 max 78.9%.

It is an Opus-class model, but faster, more token-efficient and lower cost.

SpaceXAI is also publicly conceding where it sits. Axios, summarizing the launch, reported that the company “outperforms some OpenAI and Anthropic models on speed, price and performance, but not the largest, latest models from those competitors,” and that Musk “expects to beat and exceed those models soon.” The model is served at 80 tokens per second, and the company notes the speed combined with token efficiency is what produces the lower per-task cost.

• DeepSWE 1.0 (pass@1): 62.0%
• SWE Marathon resolution rate (pass@1): 29.0%
• Served at 80 TPS
• Output tokens per SWE Bench Pro task: 15,954 (about 4.2× fewer than Opus 4.8 max)

The Trade Behind the Colossus Compute

SpaceXAI built Grok 4.5 on Colossus, the company’s GPU cluster, training across tens of thousands of NVIDIA GB300 GPUs according to the launch page. The four-line pitch SpaceX posted when it announced the April framework was that the deal couples Cursor’s leading product with “SpaceX’s million H100 equivalent Colossus training supercomputer.”

That same compute is being rented out to rivals. Axios reported on launch day that Grok 4.5 “was trained using the same compute capacity SpaceXAI is leasing to its competitors Anthropic and Google.” The Globe and Mail, working from the S-1 filing SpaceX filed around its IPO, put dollar figures on those leases: Anthropic pays SpaceX $1.25 billion per month for Colossus 1 access through May 2029; Alphabet pays $920 million per month from October 2026 through June 2029. SpaceX’s 2025 capital expenditures on AI were $12.7 billion, more than three times its space and connectivity segments combined.

• Trained on Colossus alongside Cursor’s coding data
• Same compute capacity SpaceXAI rents to Anthropic and Google
• Live in Grok Build, Cursor on all plans, and SpaceXAI console
• Not yet in the EU; mid-July EU availability expected

The trade-off is now visible on the company’s own pricing page. Cursor, the channel that converts Grok 4.5 into paying subscribers, was built on training data that lives inside a cluster that is being sold hour by hour to the labs Grok 4.5 is trying to undercut. The cross-subsidy has so far held. The question is whether it holds once Cursor’s market share stabilizes and Anthropic’s compute lease stretches out.

Frequently Asked Questions

What does Grok 4.5 cost?

Grok 4.5 lists at $2 per million input tokens and $6 per million output tokens, per the SpaceXAI launch page. Anthropic’s Opus 4.8 lists at $5 and $25, and OpenAI’s GPT-5.6 Luna lists at $1 and $6 for the same units.

Is Grok 4.5 available in Cursor?

Yes. SpaceXAI says Grok 4.5 is the default in Grok Build, available in Cursor on all plans, and accessible through the SpaceXAI console. The model is not yet available in the EU; SpaceXAI expects EU availability in mid-July.

What can Grok 4.5 do that rivals can’t?

It targets long-running engineering tasks and also legal and financial workloads, according to TNW citing Bloomberg’s coverage of the launch. The launch page shows example one-prompt apps, complex Excel models, PowerPoint diagrams, and notes for future reference.

How does Grok 4.5 compare to GPT-5.6 and Claude Opus 4.8 on benchmarks?

Grok 4.5 outperforms Opus 4.8 on DeepSWE 1.0 (62.0% vs 55.75% pass@1) and on SWE Marathon resolution rate (29.0% vs 26.0%). It trails OpenAI’s GPT-5.5 xhigh and Fable max on the same benchmarks. SpaceXAI states plainly that the model does not yet match the largest, latest models from OpenAI and Anthropic on raw performance.

Is the SpaceX-Cursor acquisition closed?

No. CNBC reported the formal agreement on June 16, 2026, and SpaceX expects the $60 billion all-stock acquisition of Cursor parent Anysphere to close during the third quarter of 2026, subject to regulatory approvals.