The Vibe Father
Head to head
Both of these are value plays, but they're value plays of very different kinds. Grok 4.5, released July 9, is the promising newcomer — a strong contest score at a low price, but with most of its slate still unpublished. DeepSeek V4 Pro is the proven value king — cheaper still, open-weight, and carrying receipts on the benchmarks that matter for building. The comparison is really promising-but-unproven versus cheap-and-proven. We run both on our live benchmarks (VCI = SWE-bench 40 / Terminal-Bench 30 / LiveCodeBench 30). Here's the honest read.
Where Grok 4.5 wins
A strong published contest score. Grok 4.5 posts 87.4 on LiveCodeBench — a genuinely strong algorithmic number that puts it right in the mix with far more expensive models on self-contained, well-specified problems. If your work leans contest-style or single-function, that's a promising signal.
A fresh flagship at a modest price. At $2 per million input and $6 per million output, Grok 4.5 is priced as a value model, not a premium one. For a brand-new flagship from a well-resourced lab, that's an aggressive price point, and the output cost in particular undercuts a lot of the field.
Room to surprise. Grok 4.5 is days old at the time of writing. Its SWE-bench and Terminal-Bench scores aren't published yet, which cuts both ways — but a strong LiveCodeBench debut is often a leading indicator, and there's genuine upside as the rest of its slate lands on independent boards.
Where DeepSeek V4 Pro wins
Proof it can do real repo work. DeepSeek V4 Pro posts 77.6 on SWE-bench Verified — the benchmark for resolving real GitHub issues across real codebases — alongside 87.5 on LiveCodeBench. Grok's SWE score is still an open question; DeepSeek's is a published fact. For building, that difference is the whole game.
Lower price and open weights. DeepSeek runs $0.435/$0.87 — a heavy month around $30, cheaper than Grok's $2/$6. And it's open-weight, so you can self-host, keep code private, and escape markup entirely. Grok is closed and, while cheap, costs more per token.
A marginally higher contest score. DeepSeek's 87.5 LiveCodeBench edges Grok's 87.4 — a wash in practice, but it means Grok isn't even winning the one axis they both publish.
The numbers side by side
| Model | SWE-bench | Terminal-Bench | LiveCodeBench | In / Out per M | Weights |
|---|---|---|---|---|---|
 Grok 4.5 | not published | not published | 87.4 | $2 / $6 | Closed |
 DeepSeek V4 Pro | 77.6 | not published | 87.5 | $0.435 / $0.87 | Open |
Reading the split correctly
Right now, this isn't close on evidence. DeepSeek publishes a repo score and Grok doesn't, DeepSeek is cheaper, and DeepSeek is open-weight where Grok is closed. On the one benchmark they both publish, they're tied. So today, DeepSeek is the stronger value play by a clear margin — it does everything Grok does on the published axis and adds a proven repo number, lower price, and open weights. The honest caveat is timing: Grok 4.5 is brand new, and its missing scores could land strong. This is a story worth re-checking as independent numbers arrive on our board — but you build with what's proven, not with what might be.
It's worth resisting the temptation to read Grok's LiveCodeBench score as a proxy for the missing ones. A strong contest number tells you a model handles self-contained algorithmic problems well; it does not tell you how the model fares on real, messy, multi-file repository work, which is a genuinely different skill. Plenty of models post excellent LiveCodeBench figures and then trail on SWE-bench — Gemini 3.1 Pro is a well-known example of exactly that gap. So Grok's 87.4 is encouraging, but it's evidence about contests, not about building, and treating it as the latter is how buyers end up disappointed. Until Grok's SWE-bench lands, the repo question is genuinely open.
There's also the openness axis, which is easy to underweight. DeepSeek being open-weight isn't only a data-control perk — it's a hedge against a vendor changing prices, deprecating a model, or throttling access. A closed model like Grok, however good and however cheap today, keeps you on someone else's roadmap. For teams that value long-run predictability, that structural difference can outweigh a small score gap in either direction, and it currently points the same way the benchmarks do: toward DeepSeek.
Who should reach for Grok 4.5
- Your work is contest-leaning and the 87.4 LiveCodeBench covers your job today.
- You want to be early on a fresh flagship and will re-evaluate as its slate fills in.
- You prefer a closed first-party model and the $2/$6 price fits your budget. More in our Grok 4.5 review.
Who should reach for DeepSeek V4 Pro
- You need proven repo competence. The 77.6 SWE-bench is real building evidence.
- You want the lowest proven cost — a ~$30 heavy month beats Grok's price too.
- You want open weights. Self-host, stay private, skip markup — see best open-weight coding models.
The honest close
Two value plays, but only one has the receipts. DeepSeek V4 Pro is the safer, cheaper, more-proven pick today — open-weight and carrying a published repo score Grok hasn't matched. Grok 4.5 is worth watching, and worth using where its contest strength fits, but its case rests on numbers that haven't landed yet. Keep both in view, route the proven builder for real work, and re-check as independent scores arrive. That per-task flexibility is what a bring-your-own-key harness like The Vibe Father is for. See our cheapest capable models roundup and the full board at /benchmarks.
