In partnership with
Reading time: 6 minutes
Greetings from above,
I want to be straight with you about something before we get into this. Grok 4.5 is a genuinely interesting model launch. But a lot of the coverage you are seeing right now is either way too hyped or way too dismissive depending on which AI camp the writer already lives in. So I went through the actual benchmark data, the pricing math, and what Elon Musk said versus what the numbers actually show — and the honest picture is more useful than either extreme.
Today we will talk about:
What Grok 4.5 actually is and what it was built specifically to do
The real benchmark story — where it wins, where it loses, and what Musk said that most people missed
Whether you should be using it, routing work to it, or completely ignoring it
Let's get into it.
Your creative brief is due Friday. Viktor wrote it Tuesday.
Tell him the campaign. Viktor pulls last quarter's performance from Meta and TikTok, scrapes competitor ads, drafts the brief, posts it for review. You edit, he ships the creative requests to your designer. Inside Slack.
What Just Happened And Why It Matters
On July 8, 2026, SpaceXAI launched Grok 4.5 — the first model release since the company went public — and its first release built specifically for coding and agentic work rather than general consumer use.
The model was trained in collaboration with Cursor, the AI coding platform SpaceX is in the process of acquiring, using debugging traces and real developer session data rather than static code repositories alone. That distinction is actually important. Most models get trained on code that exists in public repositories. Grok 4.5 got trained on how developers actually behave while writing code — the decisions, the corrections, the back and forth. That is a different kind of signal.
Musk described it on X as "an Opus-class model, but faster, more token-efficient and lower cost," referencing Anthropic's top model family. He then clarified in a follow-up: "Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster."
That second quote is the one most people skipped over. Opus 4.7 is Anthropic's previous flagship — Opus 4.8 has since succeeded it, and Anthropic's current top-of-the-line release is Claude Fable 5. So Musk's own comparison, read carefully, places Grok 4.5 a generation behind the models it is ostensibly being measured against — a fact largely lost in the framing of his public statements.
That is not a takedown. That is just the honest read of what he actually said.
HOW GROK 4.5 CAN ACTUALLY HELP YOUR BUSINESS RIGHT NOW:
At $2 per million input tokens and $6 per million output tokens, it is dramatically cheaper than Claude Opus 4.8 at $5 input and $25 output, and GPT-5.6 Sol at $5 input and $30 output — which means high-volume coding workflows that were previously expensive to run just got a lot more affordable
Grok 4.5 completes comparable coding work using 15,954 average output tokens versus 67,020 for Opus 4.8 — meaning it uses about 4.2x fewer tokens for the same task, so the cost advantage is actually bigger than the price difference alone suggests
It is already available in Grok Build, in Cursor on all plans, and via the SpaceXAI API console — so if you are already a Cursor user you have access to it today without changing anything about your setup
The Part Musk Does Not Want You To Skip Over
Ok so here is the benchmark picture laid out plainly so you can make an actual decision with it.
On Snorkel's GDPVal+ dataset — an expert-created benchmark of professional workplace reasoning tasks across the economy — Grok 4.5 achieved a mean pass rate of 29%, surpassing GPT-5.5 at 22% and Opus 4.8 at 21%. The improvement was concentrated in legal work, education and healthcare tasks — areas that demand deep professional judgment. That is a legitimately impressive result.
On DeepSWE 1.1, which scores how reliably a model closes real developer-submitted bugs, Grok 4.5 posted 53%, trailing Opus 4.8's 59% and GPT-5.5's 67%, with Claude Fable 5 topping the field at 70%.
On SWE-Bench Pro, a broader software-engineering benchmark, Grok 4.5 scored 64.7%, enough to edge out GPT-5.5's 58.6%, but still behind Opus 4.8 at 69.2% and well short of Fable 5's 80.4%.
So basically what you have is a model that beats GPT-5.5 on some coding benchmarks, loses to Opus 4.8 on the harder ones, and sits well below Fable 5 on everything raw accuracy related. xAI printed these results rather than cropping them out, which is actually worth noting. The pitch is explicitly price-performance, not supremacy.
The honest one-sentence take: Grok 4.5 benchmarks like a strong tier-two coding model, splitting results with Claude Opus 4.8 while trailing the frontier, and its standout number is output efficiency rather than any accuracy score.
The Flywheel Nobody Is Talking About
Here is the part of this story that is actually bigger than the benchmarks.
The real story is the flywheel: one owner now holds the compute, the model, and the coding tool that feeds it. SpaceX owns Colossus, the Memphis supercomputer. SpaceX is acquiring Cursor. SpaceX trained Grok 4.5 on Cursor data. And SpaceX is now using that same compute capacity that it also leases to Anthropic and Google to run its own competing models.
As its own compute needs grow, SpaceXAI may have to choose between using capacity for its own models or leasing it to others as a revenue stream. That is a very politely worded way of saying that Anthropic is currently training its models on compute owned by a direct competitor. That is a structural tension that has no clean resolution.
And then there is the speed angle that Musk flagged himself. xAI has not yet deployed its internally developed C/C++ inference stack that maps directly to the GB300 hardware, and when it does, Musk expects speeds to double or more. A further step-change release is signaled for next month.
So the model you are evaluating today is running on software that is not yet optimised for the hardware it is running on. That is actually a meaningful caveat when you are comparing latency numbers across providers.
⚙️ Here Is The Full Breakdown Of Everything New In Grok 4.5
What the model actually is under the hood
Grok 4.5 is confirmed as a mixture-of-experts system with a 500K context window, priced at $2 per million input tokens and $6 per million output tokens with a $0.50 cached input rate. It supports both the Responses API and Chat Completions formats, meaning it slots into codebases already built on OpenAI-compatible APIs without modification. If you are already calling GPT endpoints you can swap in Grok 4.5 with minimal changes.
Where it is available right now
Grok 4.5 is available today in Grok Build, in Cursor on all plans, and from the SpaceXAI API console. However Grok 4.5 is not yet available in the EU in any SpaceXAI products or the API console, with EU availability expected in mid-July 2026. If you or your team are based in Europe you are going to need to wait a couple more weeks.
The token efficiency story and why it actually matters
On SWE-Bench Pro tasks, Grok 4.5 used an average of 15,954 output tokens to complete each job while Opus 4.8 burned through 67,020 tokens for the same work — a 4.2x gap. For teams running AI at volume, that difference compounds into real savings on top of the already-lower price per token. This is probably the strongest real-world argument for the model. When you multiply cheaper tokens by fewer tokens per task the cost difference between Grok 4.5 and Opus 4.8 for high-volume coding work becomes enormous.
The speed numbers and what they mean
The model runs at 80 tokens per second, which is fast-model territory. For context most frontier reasoning models run significantly slower than this because they are spending more time thinking per token. Grok 4.5's speed is a real advantage for workflows where latency matters — live coding suggestions, real-time agent steps, anything where you are waiting on the model in a loop.
What Grok Build is adding on top
Separately from the model itself, xAI has been shipping Grok Build updates at a fairly aggressive pace. Recent additions include a new /goal slash command that lets you set a finish line for the agent rather than a single prompt, a refreshed /sessions flow that opens an Agent Dashboard, improved skill and plugin handling, and a new /docs guide picker. For developers already in the Grok Build terminal these quality of life updates are meaningful even before you factor in the model upgrade.
The one area where it clearly loses
On agent reliability across multi-step tool chains, GPT-5.6 Terra hits 97.2%, Claude Sonnet 5 scores 94.8%, and Grok 4.5 trails at 89.6%. A 7-point gap means one failed tool call every twelve agent steps — enough to break CI pipelines. If you are running complex multi-step agentic workflows where reliability is the priority, Grok 4.5 is not currently the safest default.
The practical routing strategy most teams should use
Many teams route high-volume steps to Grok 4.5 and escalate the hardest problems to a Claude model. The most cost-effective architecture in 2026 is rarely one model. This is honestly the most useful takeaway from everything that launched this week. You do not have to choose one model and commit to it. You route by task — Grok 4.5 for fast, high-volume, cost-sensitive steps, and a more capable model for the things that actually need maximum accuracy.
GROK 4.5 FULL BREAKDOWN — SUMMARY
Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, making it dramatically cheaper than both Claude Opus 4.8 and GPT-5.6 Sol — and when you combine that with 4.2x fewer output tokens per task the real-world cost difference is significantly larger than the sticker price suggests
On benchmark accuracy Grok 4.5 splits results with Opus 4.8 while trailing Claude Fable 5 on every coding metric xAI chose to publish — the pitch is price-performance not supremacy and the benchmarks xAI published themselves confirm that framing
The most effective approach is routing by job — Grok 4.5 as a cost-efficient default tier for high-volume work, with a premium Claude model reserved for the hardest reasoning tasks that genuinely need the accuracy
WRAP UP
What you learned today:
Grok 4.5 launched July 8 as xAI's first model specifically built for coding and agentic work, trained on real Cursor developer session data and priced at a level that makes it genuinely competitive for high-volume pipelines even though its raw accuracy trails Anthropic's top models
The efficiency story is the real argument — 4.2x fewer output tokens per coding task combined with already-lower per-token pricing means the total cost of running Grok 4.5 at volume can undercut rivals by a margin that is much wider than the headline price comparison suggests
The honest routing strategy right now is to use Grok 4.5 for speed and volume, escalate to Claude Opus or Fable 5 when accuracy is the priority, and stop treating any single model as the right choice for every single task
The most interesting AI model right now is not always the most capable one. Sometimes it is the one that does 80% of the work at 25% of the cost and leaves your budget intact for the 20% of tasks that actually need the frontier. Grok 4.5 is that model for a lot of coding workflows — and knowing exactly where it fits is more valuable than any benchmark number alone.
And as always, thanks for being a part of my lovely community,
Keep learning,
🔑 Robert from God of Prompt
