Grok 4.5: the price shot heard round the labs.

On July 8, SpaceXAI launched Grok 4.5 — its strongest model, built for coding, agentic tasks, and knowledge work, and trained alongside Cursor. The headline isn't a benchmark crown. It's the invoice: $2 per million input tokens, $6 per million output, served at fast-model speeds. Here's what shipped, what SpaceXAI's own charts actually show, and where you can (and can't) use it today. Updated September 22, 2026: Grok 4.6 landed on August 12 at the same $2/$6 price (section 04), and Grok 4.7 followed on September 21 — bigger base model, price still unchanged — section 05.

01 What launched

02 The honesty layer: read their own charts

SpaceXAI calls Grok 4.5 its smartest model — and Elon Musk calls it "Opus-class." Both can be true while the fine print stays interesting. On the benchmark charts in SpaceXAI's own announcement:

• Anthropic's Claude Fable 5 (max) still leads on DeepSWE 1.0 and 1.1, Terminal Bench 2.1, and SWE Bench Pro — with Grok 4.5 close behind, ahead of Opus 4.8 on several.
• Grok 4.5 takes first on SWE Marathon (long-haul software tasks), beating both Opus 4.8 and Fable.
• And July 9 muddied the water further: OpenAI's GPT-5.6 family launched the next day with its own state-of-the-art claims.

03 Where you can use it today

Three doors, all open now — plus one closed one:

• Grok Build — Grok 4.5 is now the default model, and SpaceXAI is offering free Grok 4.5 usage in Build for a limited time. It handles more than code: multi-sheet Excel models with web research, PowerPoint decks with native shapes, and Word drafting via the Office add-ins (see our Grok in Office lesson).
• Cursor — available on all Cursor plans, also with free usage for a limited time.
• The API — model ID grok-4.5 from the SpaceXAI console.
• The EU caveat — at launch Grok 4.5 was not available in the EU in any SpaceXAI product or the API, and SpaceXAI pointed at mid-July 2026. That window has since passed; we have not independently confirmed current EU status, so check the model picker in your own region rather than trusting a launch-day note.

04 August 12, 2026: Grok 4.6 lands — same price

Five weeks after 4.5, SpaceXAI and Cursor released Grok 4.6 together. The pitch is narrower and more useful than "smartest model": it is built to stay on a job across many steps — long research runs, working across a codebase, turning a rough product idea into something that actually opens in a browser.

Two things changed under the hood that you can actually feel. Cursor reports 4.6 does more self-checking on long runs — testing its own work before moving on — and produces stronger first passes on visual and interactive projects, establishing structure and visual language in one go instead of three.

Should you switch? If you are running agents that work for minutes rather than seconds, 4.6 is the one to test — that is precisely what it was trained for. For short one-shot prompts, 4.5 at the same price is unlikely to feel different, and the model ID you already have keeps working. See what Grok actually costs for how these rates sit against the subscription tiers.

05 September 21, 2026: Grok 4.7 — bigger base model, third time at $2/$6

Six weeks after 4.6, SpaceXAI shipped Grok 4.7, calling it "our most capable model for coding and knowledge work." Under the hood it is a new, larger base model (4.6 was a retrain of the 4.5 base), trained with a longer reinforcement-learning run on a harder task mix weighted toward jobs that take many hours. SpaceXAI says it verifies its own work more, handles longer context better, and was trained to natively understand the Grok Bot harness — so it should feel better in conversational and knowledge work, not only in code.

What SpaceXAI's own table shows, read against the other columns: on its headline coding benchmark (CursorBench 4.0) 4.7 scores 46.3% vs 40.4% for 4.6 — a real step — but Fable 5.1 still leads at 51.8%, and on Terminal-Bench 4.0 the gap is wider (38.0% vs 57.9%). Where 4.7 posts outright wins are narrower, vendor-picked tests: an electrical-engineering benchmark and a legal-agent benchmark where every rival scores badly. On the broad GDPval knowledge-work index it lands second (1,695 Elo) behind Fable 5.1 (1,735) and ahead of GPT-6 Astra.

Should you switch? If you already run 4.6 in an agent or coding loop, change the model ID and re-run your own eval — same price, so the only cost is the test. For Grok Build users the switch already happened: 4.7 is the new default. For short chat prompts, wait for the consumer app to say what it is serving.