The Vibe Father
Model Review
xAI's Grok 4.5 landed on July 9, 2026, and this is our fuller pass on it after living with the model for a bit. The short version has not changed since our first look: it is a genuinely strong contest coder at a friendly price, and its ability to do real repository work is still an open question because the labs have not published the numbers that would answer it. That gap is the spine of this review, so we are going to be disciplined about it.
The one benchmark we do have is 87.4 on LiveCodeBench, which slots Grok 4.5 into the frontier pack on contest-style coding. Pricing is $2 per million input tokens and $6 per million output, it streams at 91 tokens per second, and it ships a 500k-token context window. On paper that is an attractive package. The trouble is that "on paper" is doing a lot of work in that sentence.
Contest coding is not repo work
Our index at /benchmarks weights three benchmarks because they measure three different skills, and Grok 4.5 currently has a score for exactly one of them.
- LiveCodeBench measures algorithmic, contest-style problems — self-contained puzzles with clean specs and checkable answers. It is a real signal for raw coding intelligence, and Grok 4.5's 87.4 is a real result.
- SWE-bench Verified measures whether a model can resolve actual GitHub issues in real repositories — reading code it did not write, understanding context, fixing a bug without breaking the neighborhood. xAI has not yet published a Grok 4.5 SWE-bench Verified number.
- Terminal-Bench measures agentic shell work — operating a real terminal over many steps and recovering when commands fail. Also not yet published for Grok 4.5.
These skills correlate, but they do not travel together. We have watched models ace self-contained problems and then stall the instant a task requires navigating a 200-file codebase or holding a stateful multi-step shell session. That is precisely why our index weights SWE-bench Verified at 40% — repo work is the actual job — with Terminal-Bench and LiveCodeBench at 30% each. Grok 4.5 scores on the 30% we care about least for daily engineering, and is silent on the 70% we care about most.
So the honest claim is narrow: Grok 4.5 is a frontier-class contest coder at a bargain price, and we do not yet know whether it can do real repo work. Not "it can't" — we don't know. Anyone telling you otherwise before the SWE-bench and Terminal-Bench numbers land is guessing.
Where the one number lands
Here is how the published LiveCodeBench score and pricing sit against the models around it on our board.
| Model | LiveCodeBench | SWE-bench | Price (in/out per M) |
|---|---|---|---|
 Grok 4.5 | 87.4 | not yet published | $2 / $6 |
 Claude Opus 4.8 | 87.8 | 88.6 | $5 / $25 |
 Gemini 3.5 Flash | 87.6 | 79.3 | $1.50 / $9 |
 DeepSeek V4 Pro | 87.5 | 77.6 | $0.435 / $0.87 |
Two things jump out. First, the LiveCodeBench pack is tight — four models inside half a point — which means that benchmark alone cannot separate them, and that is exactly why the missing scores matter so much. The models Grok is tied with all have published SWE-bench numbers; Grok does not. Second, the pricing is aggressive: its $6 output is roughly a quarter of Opus 4.8's $25. DeepSeek V4 Pro still owns the absolute price floor, but Grok undercuts the Western flagships hard. If — big if — its repo numbers arrive anywhere near its contest number, that combination becomes one of the more interesting value plays on the board.
How to trial it without getting burned
We do not wait for labs to publish before forming a working opinion — we slot new models into low-risk seats and let real work grade them. Here is the plan we would run for Grok 4.5, and it is a template you can reuse for any release.
- Reviewer seat first. Reviewing is the ideal audition: it is cheap (mostly input tokens, and $2/M input is friendly), it is low-risk (a bad review wastes minutes; a bad merge wastes days), and it stress-tests the exact skill LiveCodeBench does not — reading unfamiliar code in context and reasoning about what is wrong. The 500k window holds a serious diff plus the surrounding files. Sharp review notes are early evidence the repo-comprehension gap may not be a gap.
- Then Builder, on a branch, behind a verification gate. If it survives the audition, it gets a real ticket on an isolated branch, with your actual build and test suite gating "done." No self-reported success. This is where "contest coder" and "repo worker" definitively part ways.
- Track the boring numbers. Tokens per completed task, retries, files touched versus files that needed touching. A model 30% cheaper per token that needs two extra rounds is not cheaper.
Verdict
Grok 4.5 is promising and unproven. The LiveCodeBench score earns it a seat in the frontier conversation, the pricing earns it a seat in the value conversation, and the 500k context is enough for serious work. What it has not earned is trust with your repository, because nobody has published the evidence yet. When SWE-bench Verified or Terminal-Bench numbers appear, they go on the leaderboard and this review gets a real verdict. Until then, audition it in the cheap seats and gate everything.
