What happened
On July 8, 2026, SpaceXAI (the division that absorbed Elon Musk's xAI in February 2026) released Grok 4.5. Musk called it "an Opus-class model, only faster, more token-efficient and cheaper," adding that by internal evaluation it is "roughly at the level of Opus 4.7, but noticeably faster" (TechCrunch).
The essentials:
- Price: $2 per million input tokens and $6 per million output — roughly half what competitors charge for their top models.
- Architecture: according to trade press, a 1.5T-parameter MoE (the V9 foundation model) with a 500,000-token context window.
- Training: on top of the usual corpus, post-training (SFT + RL) added real developer sessions from the Cursor editor — debugging traces, multi-file diffs and user edits. Earlier, on June 16, SpaceX announced the acquisition of Anysphere (Cursor's parent company) for $60B in stock.
- Availability: Grok Build, Cursor (on every plan) and the SpaceXAI console; in the EU from mid-July.
On benchmarks published by xAI itself, Grok 4.5 beats Opus 4.8 on DeepSWE 1.0 (62.0% vs 55.75%) and Terminal-Bench 2.1 (83.3% vs 78.9%), but trails on SWE-Bench Pro (64.7% vs 69.2%). The headline advantage is token economy: on SWE-Bench Pro tasks the model spends an average of 15,954 output tokens versus 67,020 for Opus 4.8 — roughly four times fewer (MarkTechPost).
Why it matters
- Price as a weapon. $2/$6 is about half of what rivals charge for frontier models. VentureBeat writes that the launch "could rattle Anthropic and OpenAI."
- A closed data flywheel. The model was post-trained on data from a live code editor that now belongs to the same corporation. This is the first high-profile case of a coding agent's data flywheel closing inside a single company.
- Careful with the numbers. The benchmarks come from xAI itself. Independent Artificial Analysis puts Grok 4.5 only fourth on its Intelligence Index (behind Claude Fable 5, GPT-5.5 and Opus 4.8), while acknowledging the best result on agentic tool use. TechTimes notes a higher hallucination rate.
- The privacy question. Per press reports, Cursor user sessions went into training — debugging, refactoring, architecture discussions. Analysts are already asking whether Cursor will stay multi-model (routing to Claude and GPT) once the deal closes.
What it means for vibe coding
- A cheap, fast default for agents. At $2/$6 and a 4x saving on output tokens, Grok 4.5 can sharply cut the cost of long agentic runs — multi-file edits, autofixes, "run it → fix it" loops. TechTimes estimates coding-agent costs drop by roughly 80%.
- Already in Cursor. The model is available in Cursor on every plan, so you can try it in your usual editor without migrating anywhere.
- Test it on your own tasks. Cheap and fast is not the same as good: Opus 4.8 still leads on SWE-Bench Pro, and independent rankings place Grok 4.5 below the top. For critical code, run models side by side.
- Think about your data. If you work with sensitive code in Cursor, re-check your privacy settings and the training policy — especially since editor data has already been used to train a model.
Sources
- TechCrunch — SpaceXAI releases Grok 4.5, which Elon describes as an «Opus-class» model
- MarkTechPost — SpaceXAI Releases Grok 4.5, a Cursor-Trained Model at $2/M Input
- VentureBeat — SpaceX's Grok 4.5 launches at half the price of rivals
- TechTimes — Grok 4.5 Cuts Coding-Agent Cost 80%: Near-Frontier Speed, Higher Hallucinations