Built with the AI editor Cursor, Grok 4.5 targets software engineers, not SWE Bench Pro, the AI coding benchmark that an OpenAI evals team member now calls 'saturated.'
xAI released Grok 4.5 on Tuesday, framing the new model as the first it has trained explicitly for software engineering and "agents," autonomous systems that take multi-step actions like opening files, running tests, and editing code rather than only answering chat prompts. The company is selling the launch on speed and price, not on topping the long-standing coding leaderboards that have, until now, defined who gets to be called a frontier model.
An OpenAI evals team member publicly described SWE-Bench Pro, the industry-standard test for AI coding ability, as "saturated" or "terminally flawed," according to a Latent Space summary of the comment. The same digest flagged "FrontierCode" as a candidate successor benchmark. (AINews digest covering the launch). Grok 4.5 is the first major model release that explicitly declines to chase that score, and that posture is the news more than the model itself.
xAI's announcement positioned the model around "capability-per-dollar" rather than raw benchmark supremacy. The launch post on X described Grok 4.5 as built for "high-performance coding and agentic tool use" and tied its value to speed and cost rather than to the next-point improvement on SWE-Bench Pro. (SpaceXAI announcement on X). In practice, that pitch targets a buyer who has to ship software against a budget: enterprise developers, internal tool teams, and coding-assistant vendors whose unit economics depend on the cost per solved task.
The collaboration with Cursor is the operational story underneath. Cursor, the AI-powered code editor that professional developers use as a daily driver, said in its own blog post that it helped train Grok 4.5 and called it "the first model we've helped build outside of pure software engineering." (Cursor blog: Introducing Grok 4.5). Cursor is a category leader in AI coding tooling, and its involvement gives xAI something hard to fake: a real, working developer toolchain, the telemetry from millions of coding sessions, and the feedback loops that come with both.
That relationship is also a corporate one. Business Insider reported in June 2026 that SpaceX is acquiring Cursor for $60 billion, a deal that would fold the editor into Musk's broader commercial empire. (Business Insider: SpaceX is buying AI coding startup Cursor for $60 billion). The "post-Cursor acquisition" framing in xAI's launch narrative leans on a deal whose closure the public receipts do not independently confirm. The operational collaboration is on the record either way, and the resulting model has Cursor's fingerprints on its training and evaluation.
On substance, xAI is positioning Grok 4.5 against Anthropic's Claude Opus, the family that has set the bar for agentic coding work over the last year. Elon Musk called Grok 4.5 "Opus-class" in comments reported by TechCrunch and Axios, a label that is Musk's framing rather than an independent benchmark categorization, and that should be read as positioned rather than proven. (TechCrunch; Axios scoop). The same coverage noted that Grok 4.5 sits in a different weight class from xAI's prior Composer series, which Latent Space puts at around 1.5 trillion parameters. That figure originates with the newsletter's synthesis rather than a direct xAI disclosure and should be read as a ballpark.
The timing is also deliberate. Grok 4.5 lands roughly a day before OpenAI's reportedly scheduled GPT 5.6 reveal, which gives xAI a narrow attention window to define the conversation before the next frontier model steals it. (AINews digest). xAI's bet is that, by the time GPT 5.6 lands, the relevant question is no longer "which model tops SWE-Bench Pro" but "which model writes and ships code fastest, cheapest, and most reliably."
What to watch: whether any successor benchmark (FrontierCode or otherwise) gains traction in the next quarter. The durability of xAI's "speed and price" pitch depends on whether the field replaces the saturated test with a new one, and on whether that replacement rewards the same kind of agentic, multi-step work that Cursor and xAI say Grok 4.5 was trained for.