Scale · founder · 7 min read

Grok 4.5 is here: an 'Opus-class' coding model, half the price

xAI shipped Grok 4.5 on July 8 — trained on real Cursor sessions, priced 60% below Opus 4.8, and live in Cursor on every plan. The founder's read.

Published July 8, 2026 ·

xai grok grok-4.5 cursor models vibe coding

xAI released Grok 4.5 on July 8, 2026 — its first model built specifically for coding and agentic work, and the one Elon Musk is calling “Opus-class.” That’s a loaded claim, so let’s be careful with it. But the more interesting detail isn’t the benchmark bragging. It’s that Grok 4.5 was trained partly on real Cursor developer sessions, and it went live inside Cursor on every plan the same day it launched. If you build with Cursor, a new frontier model just quietly appeared in your model picker.

Here’s the useful version for founders and PMs who don’t read model cards for sport: what shipped, whether the “Opus-class” line holds up, where it shows up in your stack, and what to actually do about it.

What actually shipped

Grok 4.5 is built on xAI’s 1.5-trillion-parameter V9 foundation and trained on real-world programming data — production pull requests and, notably, anonymized Cursor coding sessions that came with the SpaceX–Cursor deal earlier this year. That’s the first tangible product of that acquisition: xAI now has a pipeline of real developer behavior to train on, and Grok 4.5 is what came out the other end.

Two numbers matter. On the Artificial Analysis Intelligence Index — a broad, independent aggregate — Grok 4.5 lands fourth, above every open-weight model and above all of Google’s Gemini models. On coding specifically, independent benchmarks put it roughly on par with GPT-5.5 running in Codex, at about half the per-task cost.

The pricing is the headline for anyone paying real bills: $2 per million input tokens, $6 per million output. That’s over 60% cheaper than Claude Opus 4.8 or GPT-5.5. And it’s aggressively token-efficient — xAI clocks it at about 14,000 output tokens per Intelligence Index task, versus roughly 67,000 for Opus 4.8. Fewer tokens to reach the same answer means the real-world cost gap is even wider than the sticker price suggests.

Does “Opus-class” hold up?

Partly. Musk’s framing is that Grok 4.5 belongs in the same conversation as Anthropic’s top model, and on aggregate intelligence scores and price-efficiency, that’s defensible — it clearly outruns the mid-tier field. But “belongs in the conversation” is not the same as “wins the conversation.”

The honest read: Grok 4.5 is a genuinely strong, unusually cheap coding model that’s competitive with GPT-5.5-tier work. What it hasn’t proven yet is the thing Opus 4.8 and Fable 5 are actually known for — careful, reliable work inside large, messy, real codebases over long sessions. The benchmarks it’s winning are xAI’s own framing of “what an AI can do inside a real codebase.” Independent, adversarial testing on messy production repos takes weeks to shake out. Treat the “Opus-class” label as a marketing claim with real substance behind it, not a settled fact.

One thing genuinely in its favor: because it was trained on Cursor session data, it should feel unusually well-suited to the Cursor workflow specifically. That’s a narrower, more credible claim than “beats Opus everywhere.”

Where it shows up in the tools you use

If you use Cursor, Grok 4.5 is already in your model picker — it launched there on all plans on day one, which is unusually fast and tells you how tight the xAI–Cursor relationship now is post-acquisition. This is the most direct impact for our audience: a strong, cheap model you can select today without changing tools.

If you use Grok Build — xAI’s agentic coding CLI — Grok 4.5 is the new engine, replacing the older grok-code-fast-1 as the flagship option. See our Grok Build review for whether that terminal-first tool is even the right shape for you (for most non-technical founders, it isn’t).

If you build with Lovable, Bolt, Replit, or Base44, you won’t see “Grok 4.5” as a setting anywhere. These tools make their own routing decisions and don’t expose them. Whether any of them adopt a Musk-owned model is partly a business and trust question, not just a technical one — so don’t assume it’ll show up quietly the way an OpenAI or Anthropic model would.

One real limitation: Grok 4.5 is not yet available in the EU. xAI is targeting mid-July 2026 for EU access. If you or your users are in Europe, it may not be selectable for you yet.

What it changes about your stack

For most non-technical builders, Grok 4.5 doesn’t change which tool you use. It changes which model you can pick inside Cursor, and it pushes on the cost math in a useful direction.

The pattern here is the one this site keeps flagging: strong models are getting cheap fast. Grok 4.5 is the third aggressively-priced capable model in about a month — after GLM-5.2 and Claude Sonnet 5 — and now GPT-5.6 Terra sits in the same cheap-but-good bracket. The floor for “good enough coding model” keeps dropping. That’s straightforwardly good for founders: the tools that route to these models can offer more capable agents without raising prices, and your per-build cost trends down.

The trust dimension is the part worth naming out loud. Grok 4.5 sits inside the Musk/xAI/SpaceX orbit — the same orbit that now owns Cursor. If you’re building a customer-data-heavy app and you care about who controls your model supply chain, that concentration is a legitimate thing to weigh, the same way you’d weigh any vendor-lock question. Cursor’s bring-your-own-API-key option remains your best hedge if you want to keep model choice in your own hands. Our SpaceX–Cursor explainer covers that ownership question in more depth.

What to do this week

Three moves.

One: if you’re in Cursor, try Grok 4.5 on a real task and compare. It’s cheap, it’s fast, and it’s supposedly tuned to the Cursor workflow. Point it at something you’d normally hand to GPT-5.6 or Claude and see whether the output holds up on your actual code — not on a benchmark.

Two: don’t switch tools over it. A new model in the picker is not a reason to move off Lovable, Bolt, or Replit. If Grok 4.5 is as good as claimed, the tools that route to it will show it in the output within a couple of weeks. Let that make the case, not the launch tweet.

Three: keep your review discipline. A cheaper model tempts you to run more, review less. Resist that. Whatever model you pick, the failure modes of AI-generated code — security gaps, silent bugs, code you can’t maintain — don’t go away because the tokens got cheaper. If anything, more output means more to check.

The bigger picture hasn’t changed. Frontier-class models now ship every couple of weeks and the price keeps falling. Grok 4.5 is a real one — cheap, fast, and genuinely capable — but it’s still weather, not climate. Pick tools that absorb whichever model is best this month, keep your hand on the review process, and don’t confuse a strong launch with a settled winner.