SpaceXAI (the rebranded xAI) released Grok 4.5, a 1.5-trillion-parameter model priced at $2 per million input tokens and $6 per million output tokens. That pricing lands at roughly a third of what OpenAI and Anthropic charge for their flagship models, and the company pairs it with a claim of comparable coding quality. The public release is scheduled for Thursday, July 9; it is already live in Grok Build, Cursor, and the API console.
Key facts
- Grok 4.5 is a new 1.5-trillion-parameter V9 foundation model — a ground-up redesign, not a fine-tune of the previous architecture.
- Priced at $2 per million input tokens and $6 per million output — about a third of comparable Western frontier pricing.
- Trained in collaboration with the Cursor coding IDE, creating a feedback loop between the editor and the model.
- Primary source: the SpaceXAI Grok 4.5 announcement. Public release: July 9, 2026.
Elon Musk introduced the model with a single loaded phrase, calling it "Opus-class" — a direct comparison to Anthropic's most capable model. That is a marketing claim, not a measured result, and it is the part worth watching skeptically. What is verifiable is the architecture and the price. The 1.5-trillion-parameter V9 model is a real generational jump from the earlier V8 line, and the aggressive pricing is published, not promised.
To understand why this matters, remember what the frontier has looked like: the best models cost the most, and everyone paid the premium because nothing cheaper matched them. Grok 4.5 is the latest entrant betting that the premium is collapsing. The independent TryAI build-off put it head-to-head against GPT-5.5 and Claude on live coding tasks. The picture was mixed rather than triumphant: Grok 4.5 served responses fast (around 110 tokens per second, at a fraction of a cent per reply) and either beat or tied its rivals on a Breakout clone and a gravity simulation, but it flubbed a 3D spinning-cube task on its first attempt. That is the honest texture of a new frontier model — genuinely strong, not flawless, and cheaper than the field.
The mechanism behind the coding focus is the Cursor partnership. By training the model inside a real coding environment, SpaceXAI built what amounts to a data flywheel: the IDE generates the exact kind of interactive, tool-using coding data the model needs, and the model in turn makes the IDE better. Think of it like a chef trained not from recipe books but from years in a working kitchen — the feedback is immediate, real, and specific to the job.
Why this matters goes beyond one model. A credible frontier entrant at a third of the price pressures every incumbent's margin at exactly the moment enterprises are already fleeing expensive Western models for cheaper alternatives — the same cost story now reshaping which models US companies actually run. If "Opus-class at $2" survives contact with the July 9 release, the ceiling price for frontier intelligence just dropped again.
The honest caveat: none of the quality claims are settled. Musk's "Opus-class" label rests on internal benchmarks, and the build-off shows a model that is competitive but beatable. The public release is the real test. Until independent developers run it against their own workloads, treat the price as the confirmed news and the quality parity as the claim under investigation.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)