Grok 4.5

What it is

Grok 4.5 is xAI’s current flagship large language model, released on 8 July 2026, and it marks a clear pivot towards coding. The company built grok 4.5 for coding, agents and knowledge work, and, unusually, trained it on real developer session data from Cursor, the coding tool xAI acquired for roughly $60bn in mid-June 2026. Since that deal, some coverage has taken to calling the combined company SpaceXAI.

Trained on real coding sessions

Training on genuine Cursor sessions is the detail worth dwelling on. Rather than rely only on public code, xAI has fed grok 4.5 records of how developers actually work inside an editor: the edits, the retries and the back-and-forth of real projects. If that data proves as useful as the company hopes, it could give grok 4.5 an edge on the practical, messy business of writing and fixing software, as opposed to solving neat benchmark puzzles. The acquisition also folds a popular coding tool into xAI’s own stack, giving it both a distribution channel and a training resource.

The Cursor deal reshapes xAI’s position as much as the model does. Paying a reported figure of around $60bn for a widely used coding tool, then drawing its data and its users into the parent company, is an aggressive way to buy relevance with developers. The rebranding that some coverage now uses, SpaceXAI, captures how tightly Elon Musk’s ventures are being drawn together in the public mind, though the exact corporate structure is beyond the scope of a model page.

Price and performance

On price and performance, xAI has pitched grok 4.5 as strong value. It costs $2 per million input tokens and $6 per million output tokens, offers configurable reasoning effort, and runs at roughly 80 tokens per second. Configurable reasoning effort means a user can dial how long the model thinks before answering, trading speed for depth, and the tokens-per-second figure is a measure of how quickly it produces text. The company cites a score of 29.0% on the SWE Marathon coding benchmark against 26.0% for Claude Opus 4.8, and a fourth place on the Artificial Analysis Intelligence Index. Musk has framed the model as “Opus-class but faster and cheaper”. Those are xAI’s own figures, and independent verification is still in progress, so they are best read as claims awaiting confirmation.

What comes next

Attention is already turning to what comes next, and the phrase “grok 5 release date” is doing brisk traffic. Grok 5, rumoured to be a six-trillion-parameter mixture-of-experts model training on the Colossus 2 supercluster in Memphis, has slipped past both its first-quarter and second-quarter 2026 windows with no confirmed date. Delays at this scale are common, and xAI has not committed publicly to a new one. In the meantime the company has widened its range in other directions: Grok Voice launched on 4 June 2026, bringing spoken interaction to the line.

What to watch

The framing around Grok 4.5 is as aggressive as the pricing. Casting a model as “Opus-class but faster and cheaper” is a direct challenge to the premium labs, inviting a straight comparison on cost per unit of capability. Whether the SWE Marathon score and the index placement survive outside testing will decide how much of that framing sticks; for now they rest on xAI’s own account. A single-point lead on one coding benchmark is a thin margin, and the sensible reading is to wait for third-party numbers before treating grok 4.5 as settled above its rivals.

Where grok 4.5 sits in the wider field is as a fast, keenly priced challenger leaning hard into coding, backed by a training-data advantage that rivals cannot easily copy. The open question is whether the benchmark claims hold up under independent testing, and whether the Cursor data pays off in everyday use rather than on leaderboards. With Grok 5 running late, grok 4.5 will carry xAI’s case for some months yet. For the wider picture, see our large language models hub and the broader AI coverage.