Grok 4.5: xAI Finally Did It

Grok 4.5 shipped this afternoon, July 8, 2026, and I am writing about it the same day. That is not how I usually do these. The GPT-5.3-Codex review took me four months, and I opened it apologizing for the lateness. This one I am writing while it is still warm, because I have been waiting for it in a specific way and I want to get the reaction down before the scoreboard arguments start.

Here is the short version, and I mean it: finally. xAI finally did it. Grok is finally good — not good-for-Grok, not funny-and-fast, but good in the way that earns a place on serious work without an asterisk.

What Grok was, and what it wasn’t

I have kept Grok around for years. Not for work. For the personality, the real-time search off X, the willingness to answer a question straight instead of routing me through three disclaimers. It was the model I opened to check what was actually happening right now, or to get a blunt take, or honestly just to laugh. Loose guardrails, quick, wired into the live feed. That was the pitch and it delivered on it.

What it was not was the model you handed a long agent loop. Previous Grok generations were clever in a burst and unreliable across a span. The 4.x line before this could start a multi-step writing or refactoring task well and then drift — lose the thread halfway, forget a constraint I set three steps ago, get confident about the wrong file. For the kind of work where an agent runs unattended and every step compounds on the last, that is disqualifying. So I did what most people did: I enjoyed Grok for what it was good at and reached for Claude or Codex when the task actually mattered. Grok was the fun one. It was not the dependable one.

I say that as history, not as a grudge. I have used the serious tools all along, and I still do. But it means that when I say this release lands, it is not a fan talking himself into it. It is someone who spent years with Grok in exactly the “great personality, wouldn’t trust it with prod” box finally taking it out.

What actually shipped

xAI is positioning Grok 4.5 as its smartest model for coding, agentic tasks, and knowledge work — the trio, not just chat. The headline detail that caught my eye is that it was trained alongside Cursor. Not “integrated with” after the fact; co-trained. xAI describes tens of thousands of NVIDIA GB300 GPUs, heavy data filtering, and large-scale asynchronous reinforcement learning across hundreds of thousands of multi-step, SWE-style tasks. Whatever else you make of the numbers, that is a training recipe aimed squarely at agent work, not at winning a chat demo.

The practical facts, the ones an operator actually routes on:

• API model id is grok-4.5. It is the default in Grok Build, available in Cursor on all plans, and in the xAI API console.
• Serving speed is claimed around 80 tokens per second — “fast-model speeds,” in their words.
• API pricing is $2 per million input tokens and $6 per million output tokens, with a claim of roughly 2× better token efficiency than comparable leaders.
• There is free, limited-time usage in Grok Build and Cursor at launch.
• It is not available in the EU yet — not in xAI products, not in the API console — with mid-July 2026 given as the expected date.

The token-efficiency claim is the one worth staring at. xAI says Grok 4.5 uses about 4.2× fewer output tokens on an average SWE-Bench Pro task than Opus 4.8 at max effort — roughly 15,954 tokens versus about 67,020. That is a vendor-reported number and I will treat it like one. But if it holds up even partway, it matters more than any single benchmark row, because output tokens are what you actually pay for and what you actually wait on. A model that finishes the same task with a quarter of the tokens is cheaper and faster at the same quality, and those are the two axes that decide whether an agent is usable in a loop or just impressive in a screenshot.

There is also the pre-launch noise to set aside. Musk’s June 28 post called it close to, perhaps exceeding, Opus on the team’s own eval — which is a vendor self-eval, not evidence. Earlier press hyped some dense “1.5T” monster; the actual launch framing is a mixture-of-experts model co-trained with Cursor, and I am not going to repeat a parameter count nobody has confirmed as if it were fact. The interesting story is in the recipe and the efficiency, not in a spec-sheet legend.

The benchmarks, honestly

Let me be careful here, the same way I was in the Codex review. The benchmark tables on the launch page are vendor-selected, and I read them as directional signal, not as a coronation.

With that said, Grok 4.5 is genuinely top-tier-adjacent. On the numbers xAI published: DeepSWE 1.0 has it at 62.0%, behind Fable max (66.1%) and GPT-5.5 xhigh (64.31%) but ahead of Opus 4.8 max (55.75%). Terminal Bench 2.1 has it at 83.3%, a hair behind Fable max (84.3%) and GPT-5.5 xhigh (83.4%) and comfortably ahead of Opus 4.8 (78.9%). SWE-Bench Pro resolve puts it at 64.7%, behind Fable max (80.4%) and Opus 4.8 (69.2%) but ahead of Opus 4.7, GLM 5.2, and GPT-5.5 xhigh. On DeepSWE 1.1 it slips to 53%, below several of the same names.

So it is not the undisputed number one. Fable max still leads several of those rows, sometimes by a wide margin. Anyone who tells you Grok 4.5 just topped every coding eval is selling you the launch tweet. The real read is quieter and, for my money, more useful: it is competitive at the top on capability, and it wins on the axes the leaderboard doesn’t foreground — token cost per finished task, serving speed, and price. That combination is a different kind of good than “highest single score,” and it is the kind that changes what you actually deploy.

Why it feels good for agent work

Capability at the top of a benchmark is table stakes now. What decides whether I put a model in an agent loop is narrower: does it call tools cleanly, does it hold intent across a dozen steps, and what does a finished task cost me in time and money. Grok 4.5 is the first Grok that clears all three.

The thing that matters here is not a benchmark. It is that the model stops fighting the operator. It stays on the constraint it was given, it does not wander off to rewrite files nobody asked about, and it comes back with tool calls that parse the first time. The token efficiency is the quiet multiplier: responses land fast at that 80-TPS clip, and a multi-step task does not silently burn a fortune while it runs unattended. That is the difference between a clever model and a model you can trust to run in a loop, and it is exactly the gap the previous Grok generations never closed.

Where Claude Code still wins

Grok being good is not the same as Grok replacing everything, and it is worth being honest about the split. A model clearing the top of the coding evals does not retire the coding agent you already have working. Claude Code — like the other terminal coding agents — earns its keep on delegated coding: “go touch these files, under these rules, in this repo,” running as an isolated agent with its own scoped access. That is a workflow worth trusting, and the boundary between a model and your systems is the part that actually matters. Grok 4.5 being good enough to lead the reasoning, the planning, and the fast back-and-forth is not a reason to throw out a tool that already does its job. Multi-model is not a hedge to be embarrassed about. It is the correct posture when the models are this close and the switching cost of a setup that already works is real.

The knowledge-work side

One more thing worth a paragraph, even on a DevOps blog. xAI is pushing Grok 4.5 hard on office and knowledge work: Excel workbooks with multi-sheet models built from web research and its own notes, PowerPoint decks with native shapes, Word prose, and plugins into the Office suite. There are the usual one-prompt-app demos too — a Three.js solar system and the like — which I file under marketing until I see independent proof. I am not going to pretend I have stress-tested the spreadsheet story. But it signals where this is aimed: not just at the terminal, at the whole desk. That expansion is an interesting direction for the product, even if it is not the reason the release impresses.

Caveats, because there are always caveats

The numbers on the launch page are xAI’s numbers. Treat the coding scores, the token-efficiency multiplier, and the pricing-efficiency claim as vendor-reported until someone independent reproduces them. Verify pricing and the API docs yourself before you route production traffic — I have seen too many “reported” figures on cached input and context tiers that I could not confirm on the official page, and I am not going to repeat them here as fact.

The EU delay is a real operational footnote. If your team or your users sit inside the EU, grok-4.5 is not something you can lean on today; it is a mid-July maybe. Plan around that, don’t assume it.

And the structural one, the concern that outlasts the launch: this is a lot of concentration in one place. The compute, the model, and the Cursor tooling-and-data pipeline are increasingly the same story from the same company. I am not reaching for a conspiracy — co-training with Cursor is plausibly why the model is this good at agent work. But when one vendor owns that much of the stack, “cheaper and faster” can quietly become “harder to leave,” and that is a thing to notice while you are enjoying the cheaper and faster. The lesson from every recent frontier episode is the same one: keep model selection configurable and keep a fallback you have actually run.

Where I land

Grok is finally good. xAI finally did it. After years of a model I liked but wouldn’t trust past a single clever answer, 4.5 is the one that earns a place in the middle of real work — competitive at the top of the coding evals, ahead of the field on the cost-and-speed axes that actually decide deployment, and stable enough across an agent loop to stop being the model you babysit.

None of that retires the tools already doing their job, and it does not erase the caveats: verify the vendor numbers, watch the EU date, and don’t hand any single company the whole stack without noticing you did it. But the headline holds. This one lands. For the first time, reaching for Grok on serious work is a rational default and not a contrarian bit.